fresh start. Another try

2026-04-27 16:46:36 +00:00 · 2026-04-25 23:13:40 +03:00
parent 17d7730048
commit 2178737b36
101 changed files with 1 additions and 15518 deletions
@@ -1,91 +0,0 @@
-# Question Decomposition — Solution Assessment (Mode B, Draft02)
-
-## Original Question
-Assess solution_draft02.md for weak points, security vulnerabilities, and performance bottlenecks, then produce a revised solution draft03.
-
-## Active Mode
-Mode B: Solution Assessment — `solution_draft02.md` is the highest-numbered draft.
-
-## Question Type Classification
- **Primary**: Problem Diagnosis — identify weak points, vulnerabilities, bottlenecks in draft02
- **Secondary**: Decision Support — evaluate alternatives for identified issues
-
-## Research Subject Boundary Definition
-
-| Dimension | Boundary |
-|-----------|----------|
-| **Domain** | GPS-denied UAV visual navigation, aerial geo-referencing |
-| **Geography** | Eastern/southern Ukraine (left of Dnipro River) — steppe terrain |
-| **Hardware** | Desktop/laptop with NVIDIA RTX 2060+, 16GB RAM, 6GB VRAM |
-| **Software** | Python ecosystem, GPU-accelerated CV/ML |
-| **Timeframe** | Current state-of-the-art (2024-2026), production-ready tools |
-| **Scale** | 500-3000 images per flight, up to 6252×4168 resolution |
-
-## Problem Context Summary
- UAV aerial photos taken consecutively ~100m apart, downward non-stabilized camera
- Only starting GPS known — must determine GPS for all subsequent images
- Must handle: sharp turns, outlier photos (up to 350m gap), disconnected route segments
- Processing <5s/image, real-time SSE streaming, REST API service
- No IMU data available
- Camera: 26MP (6252×4168), 25mm focal length, 23.5mm sensor width, 400m altitude
-
-## Decomposed Sub-Questions
-
-### A: DINOv2 Cross-View Retrieval Viability
-"Is DINOv2 proven for UAV-to-satellite coarse retrieval? What are real-world performance numbers? What search radius is realistic?"
-
-### B: XFeat Reliability for Aerial VO
-"Is XFeat proven for aerial visual odometry? How does it compare to SuperPoint in aerial scenes specifically? What are known failure modes?"
-
-### C: LightGlue ONNX on RTX 2060 (Turing)
-"Does LightGlue-ONNX work reliably on Turing architecture? What precision (FP16/FP32)? What are actual benchmarks?"
-
-### D: GTSAM iSAM2 Factor Graph Design
-"Is the proposed factor graph structure sound? Are the noise models appropriate? Are custom factors (DEM, drift limit) well-specified?"
-
-### E: Copernicus DEM Integration
-"How is Copernicus DEM accessed programmatically? Is it truly free? What are the actual API requirements?"
-
-### F: Homography Decomposition Robustness
-"How reliable is cv2.decomposeHomographyMat selection heuristic when UAV changes direction? What are failure modes?"
-
-### G: Image Rotation Handling Completeness
-"Is heading-based rotation normalization sufficient? What if heading estimate is wrong early in a segment?"
-
-### H: Memory Model Under Load
-"Can DINOv2 embeddings + SuperPoint features + GTSAM factor graph + satellite cache fit within 16GB RAM and 6GB VRAM during a 3000-image flight?"
-
-### I: Satellite Match Failure Cascading
-"What happens when satellite matching fails for 50+ consecutive frames? How does the 100m drift limit interact with extended VO-only sections?"
-
-### J: Multi-Provider Tile Schema Compatibility
-"Do Google Maps and Mapbox use the same tile coordinate system? What are the practical differences in switching providers?"
-
-### K: Security Attack Surface
-"What are the remaining security vulnerabilities beyond JWT auth? SSE connection abuse? Image processing exploits?"
-
-### L: Recent Advances (2025-2026)
-"Are there newer models or approaches published since draft02 that could improve accuracy or performance?"
-
-### M: End-to-End Processing Time Budget
-"Is the total per-frame time budget realistic when all components run together? What is the critical path?"
-
---
-
-## Timeliness Sensitivity Assessment
-
- **Research Topic**: GPS-denied UAV visual navigation — assessment of solution_draft02 architecture and component choices
- **Sensitivity Level**: 🟠 High
- **Rationale**: CV feature matching models (SuperPoint, LightGlue, XFeat, DINOv2) evolve rapidly with new versions and competitors. GTSAM is stable. Satellite tile API pricing/limits change. Core algorithms (homography, VO) are stable.
- **Source Time Window**: 12 months (2025-2026)
- **Priority official sources to consult**:
-  1. GTSAM official documentation and PyPI (factor type compatibility)
-  2. LightGlue-ONNX GitHub (Turing GPU compatibility)
-  3. Google Maps Tiles API documentation (pricing, session tokens)
-  4. DINOv2 official repo (model variants, VRAM)
-  5. faiss wiki (GPU memory allocation)
- **Key version information to verify**:
-  - GTSAM: 4.2 stable, 4.3 alpha (breaking changes)
-  - LightGlue-ONNX: FP16 on Turing, FP8 requires Ada Lovelace
-  - Pillow: ≥11.3.0 required (CVE-2025-48379)
-  - FastAPI: ≥0.135.0 (SSE support)
@@ -1,212 +0,0 @@
-# Source Registry — Draft02 Assessment
-
-## Source #1
- **Title**: GTSAM GPSFactor Class Reference
- **Link**: https://gtsam.org/doxygen/a04084.html
- **Tier**: L1
- **Publication Date**: 2025 (latest docs)
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: GTSAM 4.2
- **Target Audience**: GTSAM users building factor graphs with GPS constraints
- **Research Boundary Match**: ✅ Full match
- **Summary**: GPSFactor and GPSFactor2 work with Pose3/NavState, NOT Pose2. For 2D position constraints, PriorFactorPoint2 or custom factors are needed.
- **Related Sub-question**: D (GTSAM iSAM2 Factor Graph Design)
-
-## Source #2
- **Title**: GTSAM Pose2 SLAM Example
- **Link**: https://gtbook.github.io/gtsam-examples/Pose2SLAMExample.html
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: GTSAM 4.2
- **Target Audience**: GTSAM users
- **Research Boundary Match**: ✅ Full match
- **Summary**: BetweenFactorPose2 provides odometry constraints with noise model Diagonal.Sigmas(Point3(sigma_x, sigma_y, sigma_theta)). PriorFactorPose2 anchors poses.
- **Related Sub-question**: D
-
-## Source #3
- **Title**: GTSAM Python pip install version compatibility (PyPI)
- **Link**: https://pypi.org/project/gtsam/
- **Tier**: L1
- **Publication Date**: 2026-01
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: gtsam 4.2 (stable), gtsam-develop 4.3a1 (alpha)
- **Target Audience**: Python developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: GTSAM 4.2 stable on pip. 4.3 alpha has breaking changes (C++17, Boost removal). Known issues with Eigen 5.0.0, ARM64 builds. Stick with 4.2 for production.
- **Related Sub-question**: D
-
-## Source #4
- **Title**: LightGlue-ONNX repository
- **Link**: https://github.com/fabio-sim/LightGlue-ONNX
- **Tier**: L1
- **Publication Date**: 2026-01 (last updated)
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: LightGlue-ONNX (supports ONNX Runtime + TensorRT)
- **Target Audience**: Computer vision developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: ONNX export with 2-4x speedup over PyTorch. FP16 works on Turing (RTX 2060). FP8 requires Ada Lovelace/Hopper. Mixed precision supported since July 2023.
- **Related Sub-question**: C (LightGlue ONNX on RTX 2060)
-
-## Source #5
- **Title**: LightGlue rotation issue #64
- **Link**: https://github.com/cvg/LightGlue/issues/64
- **Tier**: L4
- **Publication Date**: 2023
- **Timeliness Status**: ✅ Currently valid (issue still open)
- **Target Audience**: LightGlue users
- **Research Boundary Match**: ✅ Full match
- **Summary**: SuperPoint+LightGlue not rotation-invariant. Fails at 90°/180°. Workaround: try rotating images by {0°, 90°, 180°, 270°}. Steerable CNNs proposed but not available.
- **Related Sub-question**: G (Image Rotation Handling)
-
-## Source #6
- **Title**: SIFT+LightGlue for UAV Image Mosaicking (ISPRS 2025)
- **Link**: https://isprs-archives.copernicus.org/articles/XLVIII-2-W11-2025/169/2025/
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: SIFT+LightGlue hybrid achieves robust matching in low-texture and high-rotation UAV scenarios. Outperforms both pure SIFT and SuperPoint+LightGlue.
- **Related Sub-question**: G
-
-## Source #7
- **Title**: DINOv2-Based UAV Visual Self-Localization
- **Link**: https://ui.adsabs.harvard.edu/abs/2025IRAL...10.2080Y/abstract
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV localization researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: DINOv2 with adaptive enhancement achieves 86.27 R@1 on DenseUAV benchmark for UAV-to-satellite matching. Proves DINOv2 viable for coarse retrieval.
- **Related Sub-question**: A (DINOv2 Cross-View Retrieval)
-
-## Source #8
- **Title**: SatLoc-Fusion (Remote Sensing 2025)
- **Link**: https://www.mdpi.com/2072-4292/17/17/3048
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV navigation researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Hierarchical: DINOv2 absolute + XFeat VO + optical flow. <15m error, >2Hz on 6 TFLOPS edge. Adaptive confidence-based fusion. Validates our approach architecture.
- **Related Sub-question**: A, B
-
-## Source #9
- **Title**: XFeat: Accelerated Features (CVPR 2024)
- **Link**: https://arxiv.org/abs/2404.19174
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: CV researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: 5x faster than SuperPoint. Real-time on CPU. Semi-dense matching. XFeat has built-in matcher for fast VO; also compatible with LightGlue via xfeat-lightglue models.
- **Related Sub-question**: B (XFeat Reliability)
-
-## Source #10
- **Title**: XFeat + LightGlue compatibility (GitHub issue #128)
- **Link**: https://github.com/cvg/LightGlue/issues/128
- **Tier**: L4
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: LightGlue/XFeat users
- **Research Boundary Match**: ✅ Full match
- **Summary**: XFeat-LightGlue trained models available on HuggingFace (vismatch/xfeat-lightglue). Also ONNX export available. XFeat's built-in matcher is separate.
- **Related Sub-question**: B
-
-## Source #11
- **Title**: DINOv2 VRAM usage by model variant
- **Link**: https://blog.iamfax.com/tech/image-processing/dinov2/
- **Tier**: L3
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Developers deploying DINOv2
- **Research Boundary Match**: ✅ Full match
- **Summary**: ViT-S/14: ~300MB VRAM, 0.05s/img. ViT-B/14: ~600MB, 0.1s/img. ViT-L/14: ~1.5GB, 0.35s/img. ViT-G/14: ~5GB, 2s/img.
- **Related Sub-question**: H (Memory Model)
-
-## Source #12
- **Title**: Copernicus DEM on AWS Open Data
- **Link**: https://registry.opendata.aws/copernicus-dem/
- **Tier**: L1
- **Publication Date**: Ongoing
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Developers needing DEM data
- **Research Boundary Match**: ✅ Full match
- **Summary**: Free access via S3 without authentication. Cloud Optimized GeoTIFFs, 1x1 degree tiles, 30m resolution. `aws s3 ls --no-sign-required s3://copernicus-dem-30m/`
- **Related Sub-question**: E (Copernicus DEM)
-
-## Source #13
- **Title**: Google Maps Tiles API Usage and Billing
- **Link**: https://developers.google.com/maps/documentation/tile/usage-and-billing
- **Tier**: L1
- **Publication Date**: 2026-02 (updated)
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Google Maps API consumers
- **Research Boundary Match**: ✅ Full match
- **Summary**: 100K free requests/month. 6,000/min, 15,000/day rate limits. $200 monthly credit expired Feb 2025. Requires session tokens.
- **Related Sub-question**: J
-
-## Source #14
- **Title**: Google Maps vs Mapbox tile schema
- **Link**: https://developers.google.com/maps/documentation/tile/2d-tiles-overview
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: Both use z/x/y Web Mercator tiles (256px). Compatible coordinate systems. Google requires session tokens; Mapbox requires API tokens. Mapbox global to zoom 16, regional to 21+.
- **Related Sub-question**: J
-
-## Source #15
- **Title**: FastAPI SSE connection cleanup issues (sse-starlette #99)
- **Link**: https://github.com/sysid/sse-starlette/issues/99
- **Tier**: L4
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: FastAPI SSE developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Async generators cannot be easily cancelled once awaited. Use EventPublisher pattern with asyncio.Queue for proper cleanup. Prevents shutdown hangs and connection lingering.
- **Related Sub-question**: K (Security/Stability)
-
-## Source #16
- **Title**: OpenCV decomposeHomographyMat issues (#23282)
- **Link**: https://github.com/opencv/opencv/issues/23282
- **Tier**: L4
- **Publication Date**: 2023
- **Timeliness Status**: ✅ Currently valid
- **Summary**: decomposeHomographyMat can return non-orthogonal rotation matrices. Returns 4 solutions. Positive depth constraint needed for disambiguation. Calibration matrix K precision critical.
- **Related Sub-question**: F (Homography Decomposition)
-
-## Source #17
- **Title**: CVE-2025-48379: Pillow Heap Buffer Overflow
- **Link**: https://nvd.nist.gov/vuln/detail/CVE-2025-48379
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: Pillow 11.2.0-11.2.1 affected, fixed in 11.3.0
- **Summary**: Heap buffer overflow in Pillow's image encoding. Requires pinning Pillow ≥11.3.0 and validating image formats.
- **Related Sub-question**: K (Security)
-
-## Source #18
- **Title**: SALAD: Optimal Transport Aggregation for Visual Place Recognition
- **Link**: https://arxiv.org/abs/2311.15937
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Summary**: DINOv2+SALAD outperforms NetVLAD. Single-stage retrieval, no re-ranking. 30min training. Optimal transport aggregation better than raw DINOv2 CLS token for retrieval.
- **Related Sub-question**: A
-
-## Source #19
- **Title**: NaviLoc: Trajectory-Level Visual Localization
- **Link**: https://www.mdpi.com/2504-446X/10/2/97
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: Treats VPR as noisy measurement, uses trajectory-level optimization. 19.5m MLE, 16x improvement over per-frame VPR. Validates trajectory optimization approach.
- **Related Sub-question**: L (Recent Advances)
-
-## Source #20
- **Title**: FAISS GPU memory management
- **Link**: https://github.com/facebookresearch/faiss/wiki/Faiss-on-the-GPU
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: GPU faiss allocates ~2GB scratch space by default. On 6GB VRAM RTX 2060, CPU-based faiss recommended. Supports CPU-GPU interop.
- **Related Sub-question**: H (Memory)
@@ -1,142 +0,0 @@
-# Fact Cards — Draft02 Assessment
-
-## Fact #1
- **Statement**: GTSAM `GPSFactor` works with `Pose3` variables, NOT `Pose2`. `GPSFactor2` works with `NavState`. Neither accepts `Pose2`. For 2D position constraints, use `PriorFactorPoint2` or a custom factor.
- **Source**: [Source #1] https://gtsam.org/doxygen/a04084.html, [Source #2]
- **Phase**: Assessment
- **Target Audience**: GPS-denied UAV navigation developers
- **Confidence**: ✅ High (official GTSAM documentation)
- **Related Dimension**: Factor Graph Design
-
-## Fact #2
- **Statement**: GTSAM 4.2 is stable on pip. GTSAM 4.3 alpha has breaking changes (C++17 migration, Boost removal). Known pip dependency resolution issues for Python bindings (Sept 2025). Production should use 4.2.
- **Source**: [Source #3] https://pypi.org/project/gtsam/
- **Phase**: Assessment
- **Confidence**: ✅ High (PyPI official)
- **Related Dimension**: Factor Graph Design
-
-## Fact #3
- **Statement**: LightGlue-ONNX achieves 2-4x speedup over compiled PyTorch. FP16 works on Turing (RTX 2060). FP8 requires Ada Lovelace/Hopper — falls back to higher precision on Turing. Mixed precision supported since July 2023.
- **Source**: [Source #4] https://github.com/fabio-sim/LightGlue-ONNX
- **Phase**: Assessment
- **Confidence**: ✅ High (official repo, benchmarks)
- **Related Dimension**: Processing Performance
-
-## Fact #4
- **Statement**: SuperPoint and LightGlue are NOT rotation-invariant. Performance degrades significantly at 90°/180° rotations. Practical workaround: try matching at {0°, 90°, 180°, 270°} rotations. SIFT+LightGlue hybrid is proven better for high-rotation UAV scenarios (ISPRS 2025).
- **Source**: [Source #5] issue #64, [Source #6] ISPRS 2025
- **Phase**: Assessment
- **Confidence**: ✅ High (confirmed in official issue + peer-reviewed paper)
- **Related Dimension**: Rotation Handling
-
-## Fact #5
- **Statement**: DINOv2 achieves 86.27 R@1 on DenseUAV benchmark for UAV-to-satellite matching (with adaptive enhancement). Raw DINOv2 performance is lower but still viable for coarse retrieval.
- **Source**: [Source #7] https://ui.adsabs.harvard.edu/abs/2025IRAL...10.2080Y/
- **Phase**: Assessment
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: Satellite Matching
-
-## Fact #6
- **Statement**: SatLoc-Fusion validates the exact architecture pattern: DINOv2 for absolute geo-localization + XFeat for VO + optical flow for velocity. Achieves <15m error at >2Hz on 6 TFLOPS edge hardware. Uses adaptive confidence-based fusion.
- **Source**: [Source #8] https://www.mdpi.com/2072-4292/17/17/3048
- **Phase**: Assessment
- **Confidence**: ✅ High (peer-reviewed, dataset provided)
- **Related Dimension**: Overall Architecture
-
-## Fact #7
- **Statement**: XFeat is 5x faster than SuperPoint. Runs real-time on CPU. Has built-in matcher for fast matching. Also compatible with LightGlue via xfeat-lightglue trained models (HuggingFace vismatch/xfeat-lightglue). ONNX export available.
- **Source**: [Source #9] CVPR 2024, [Source #10] GitHub
- **Phase**: Assessment
- **Confidence**: ✅ High (CVPR paper + working implementations)
- **Related Dimension**: Feature Extraction
-
-## Fact #8
- **Statement**: DINOv2 VRAM: ViT-S/14 ~300MB (0.05s/img), ViT-B/14 ~600MB (0.1s/img), ViT-L/14 ~1.5GB (0.35s/img). On GTX 1080.
- **Source**: [Source #11] blog benchmark
- **Phase**: Assessment
- **Confidence**: ⚠️ Medium (third-party benchmark, not official)
- **Related Dimension**: Memory Model
-
-## Fact #9
- **Statement**: Copernicus DEM GLO-30 is freely available on AWS S3 without authentication: `s3://copernicus-dem-30m/`. Cloud Optimized GeoTIFFs, 30m resolution, global coverage. Alternative: Sentinel Hub API (requires free registration).
- **Source**: [Source #12] AWS Registry
- **Phase**: Assessment
- **Confidence**: ✅ High (AWS official registry)
- **Related Dimension**: DEM Integration
-
-## Fact #10
- **Statement**: Google Maps Tiles API: 100K free requests/month, 15,000/day, 6,000/min. Requires session tokens (not just API key). $200 monthly credit expired Feb 2025. February 2026: split quota buckets for 2D and Street View tiles.
- **Source**: [Source #13] Google official docs
- **Phase**: Assessment
- **Confidence**: ✅ High (official documentation)
- **Related Dimension**: Satellite Provider
-
-## Fact #11
- **Statement**: Google Maps and Mapbox both use z/x/y Web Mercator tiles (256px). Compatible coordinate systems. Main differences: authentication method (session tokens vs API tokens), max zoom levels (Google: 22, Mapbox: global 16, regional 21+).
- **Source**: [Source #14] Google + Mapbox docs
- **Phase**: Assessment
- **Confidence**: ✅ High (official docs)
- **Related Dimension**: Multi-Provider Cache
-
-## Fact #12
- **Statement**: FastAPI SSE async generators cannot be easily cancelled once awaited. Causes shutdown hangs, connection lingering. Solution: EventPublisher pattern with asyncio.Queue for proper lifecycle management.
- **Source**: [Source #15] sse-starlette issue #99
- **Phase**: Assessment
- **Confidence**: ⚠️ Medium (community report, confirmed by library maintainer)
- **Related Dimension**: API Stability
-
-## Fact #13
- **Statement**: cv2.decomposeHomographyMat returns up to 4 solutions. Can return non-orthogonal rotation matrices. Disambiguation requires: positive depth constraint + calibration matrix K precision. Not just "motion consistent with previous direction."
- **Source**: [Source #16] OpenCV issue #23282
- **Phase**: Assessment
- **Confidence**: ✅ High (confirmed OpenCV issue)
- **Related Dimension**: VO Robustness
-
-## Fact #14
- **Statement**: Pillow CVE-2025-48379: heap buffer overflow in 11.2.0-11.2.1. Fixed in 11.3.0. Image processing pipeline must pin Pillow ≥11.3.0.
- **Source**: [Source #17] NVD
- **Phase**: Assessment
- **Confidence**: ✅ High (CVE database)
- **Related Dimension**: Security
-
-## Fact #15
- **Statement**: DINOv2+SALAD (optimal transport aggregation) outperforms raw DINOv2 CLS token for visual place recognition. Single-stage, no re-ranking needed. 30-minute training. Better suited for coarse retrieval than raw cosine similarity on CLS tokens.
- **Source**: [Source #18] arXiv
- **Phase**: Assessment
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: Satellite Matching
-
-## Fact #16
- **Statement**: FAISS GPU allocates ~2GB scratch space by default. On 6GB VRAM RTX 2060, GPU faiss would consume 33% of VRAM just for indexing. CPU-based faiss recommended for this hardware profile.
- **Source**: [Source #20] FAISS wiki
- **Phase**: Assessment
- **Confidence**: ✅ High (official wiki)
- **Related Dimension**: Memory Model
-
-## Fact #17
- **Statement**: NaviLoc achieves 19.5m MLE by treating VPR as noisy measurement and optimizing at trajectory level (not per-frame). 16x improvement over per-frame approaches. Validates trajectory-level optimization concept.
- **Source**: [Source #19]
- **Phase**: Assessment
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: Optimization Strategy
-
-## Fact #18
- **Statement**: Mapbox satellite imagery for Ukraine: no specific update schedule. Uses Maxar Vivid product. Coverage to zoom 16 globally (~2.5m/px), regional zoom 18+ (~0.6m/px). No guarantee of Ukraine freshness — likely 2+ years old in conflict areas.
- **Source**: [Source #14] Mapbox docs
- **Phase**: Assessment
- **Confidence**: ⚠️ Medium (general Mapbox info, no Ukraine-specific data)
- **Related Dimension**: Satellite Provider
-
-## Fact #19
- **Statement**: For 2D factor graphs, GTSAM uses Pose2 (x, y, theta) with BetweenFactorPose2 for odometry and PriorFactorPose2 for anchoring. Position-only constraints use PriorFactorPoint2. Custom factors via Python callbacks are supported but slower than C++ factors.
- **Source**: [Source #2] GTSAM by Example
- **Phase**: Assessment
- **Confidence**: ✅ High (official GTSAM examples)
- **Related Dimension**: Factor Graph Design
-
-## Fact #20
- **Statement**: XFeat's built-in matcher (match_xfeat) is fastest for VO (~15ms total extraction+matching). xfeat-lightglue is higher quality but slower. For satellite matching where accuracy matters more, SuperPoint+LightGlue remains the better choice.
- **Source**: [Source #9] CVPR 2024, [Source #10]
- **Phase**: Assessment
- **Confidence**: ⚠️ Medium (inference from multiple sources)
- **Related Dimension**: Feature Extraction Strategy
@@ -1,31 +0,0 @@
-# Comparison Framework — Draft02 Assessment
-
-## Selected Framework Type
-Problem Diagnosis + Decision Support
-
-## Selected Dimensions
-1. Factor Graph Design Correctness
-2. VRAM/Memory Budget Feasibility
-3. Rotation Handling Completeness
-4. VO Robustness (Homography Decomposition)
-5. Satellite Matching Reliability
-6. Concurrency & Pipeline Architecture
-7. Security Attack Surface
-8. API/SSE Stability
-9. Provider Integration Completeness
-10. Drift Management Strategy
-
-## Findings Matrix
-
-| Dimension | Draft02 Approach | Weak Point | Severity | Proposed Fix | Factual Basis |
-|-----------|------------------|------------|----------|--------------|---------------|
-| Factor Graph | Pose2 + GPSFactor + custom DEM/drift factors | GPSFactor requires Pose3, not Pose2. Custom Python factors are slow. | **Critical** | Use Pose2 + BetweenFactorPose2 + PriorFactorPoint2 for satellite anchors. Convert lat/lon to local ENU. Avoid Python custom factors. | Fact #1, #2, #19 |
-| VRAM Budget | DINOv2 + SuperPoint + LightGlue ONNX + faiss GPU | No model specified for DINOv2. faiss GPU uses 2GB scratch. Combined VRAM could exceed 6GB. | **High** | Use DINOv2 ViT-S/14 (300MB). faiss on CPU only. Sequence model loading (not concurrent). Explicit budget: XFeat 200MB + DINOv2-S 300MB + SuperPoint 400MB + LightGlue 500MB. | Fact #8, #16 |
-| Rotation Handling | Heading-based rectification + SIFT fallback | No heading at segment start. No trigger criteria for SIFT vs rotation retry. Multi-rotation matching not mentioned. | **High** | At segment start: try 4 rotations {0°, 90°, 180°, 270°}. After heading established: rectify. SIFT fallback when SuperPoint inlier ratio < 0.2. | Fact #4 |
-| Homography Decomposition | Motion consistency selection | Only "motion consistent with previous direction" — underspecified. 4 solutions possible. Non-orthogonal matrices can occur. | **Medium** | Positive depth constraint first. Then normal direction check (plane normal should point up). Then motion consistency. Orthogonality check on R. | Fact #13 |
-| Satellite Coarse Retrieval | DINOv2 + faiss cosine similarity | Raw DINOv2 CLS token suboptimal for retrieval. SALAD aggregation proven better. | **Medium** | Use DINOv2+SALAD or at minimum use patch-level features, not just CLS token. Alternatively, fine-tune DINOv2 on remote sensing. | Fact #5, #15 |
-| Concurrency Model | "Async — don't block VO pipeline" | No concrete concurrency design. GPU can't run two models simultaneously. | **High** | Sequential GPU: XFeat VO first (15ms), then async satellite matching on same GPU. Use asyncio for I/O (tile download, DEM fetch). CPU faiss for retrieval. | Fact #8, #16 |
-| Security | JWT + rate limiting + CORS | No image format validation. No Pillow version pinning. No SSE abuse protection beyond connection limits. No sandbox for image processing. | **Medium** | Pin Pillow ≥11.3.0. Validate image magic bytes. Limit image dimensions before loading. Memory-map large images. CSP headers. | Fact #14 |
-| SSE Stability | FastAPI EventSourceResponse | Async generator cleanup issues on shutdown. No heartbeat. No reconnection strategy. | **Medium** | Use asyncio.Queue-based EventPublisher. Add SSE heartbeat every 15s. Include Last-Event-ID for reconnection. | Fact #12 |
-| Provider Integration | Google Maps + Mapbox + user tiles | Google requires session tokens (not just API key). 15K/day limit = ~7 flights from cache misses. Mapbox Ukraine coverage uncertain. | **Medium** | Implement session token management for Google. Add Bing Maps as third provider. Document DEM+tile download budget per flight. | Fact #10, #11, #18 |
-| Drift Management | 100m cumulative drift limit factor | Custom factor. If satellite fails for 50+ frames, no anchors → drift factor has nothing to constrain against. | **High** | Add dead-reckoning confidence decay: after N frames without anchor, emit warning + request user input. Track estimated drift explicitly. Set hard limit for user input request. | Fact #17 |
@@ -1,192 +0,0 @@
-# Reasoning Chain — Draft02 Assessment
-
-## Dimension 1: Factor Graph Design Correctness
-
-### Fact Confirmation
-According to Fact #1, GTSAM's `GPSFactor` class works exclusively with `Pose3` variables. `GPSFactor2` works with `NavState`. Neither accepts `Pose2`. According to Fact #19, `PriorFactorPoint2` provides 2D position constraints, and `BetweenFactorPose2` provides 2D odometry constraints.
-
-### Problem
-Draft02 specifies `Pose2 (x, y, heading)` variables but lists `GPSFactor` for satellite anchors. This is an API mismatch — the code would fail at runtime.
-
-### Solution
-Two valid approaches:
-1. **Pose2 graph (recommended)**: Use `Pose2` variables + `BetweenFactorPose2` for VO + `PriorFactorPose2` for satellite anchors (constraining full pose when heading is available) or use a custom partial factor that constrains only the position part of Pose2. Convert WGS84 to local ENU coordinates centered on starting GPS.
-2. **Pose3 graph**: Use `Pose3` with fixed altitude. More accurate but adds unnecessary complexity for 2D problem.
-
-The custom DEM terrain factor and drift limit factor also need reconsideration: Python custom factors invoke a Python callback per optimization step, which is slow. DEM terrain is irrelevant for 2D Pose2 (altitude is not a variable). Drift should be managed by the Segment Manager logic, not as a factor.
-
-### Conclusion
-Switch to Pose2 + BetweenFactorPose2 + PriorFactorPose2 (or partial position prior). Remove DEM terrain factor (handle elevation in GSD calculation outside the graph). Remove drift limit factor (handle in Segment Manager). This simplifies the factor graph and avoids Python callback overhead.
-
-### Confidence
-✅ High — based on official GTSAM documentation
-
---
-
-## Dimension 2: VRAM/Memory Budget Feasibility
-
-### Fact Confirmation
-According to Fact #8: DINOv2 ViT-S/14 ~300MB, ViT-B/14 ~600MB. Fact #16: faiss GPU uses ~2GB scratch. Fact #3: LightGlue ONNX FP16 works on RTX 2060.
-
-### Problem
-Draft02 doesn't specify DINOv2 model size. Proposing faiss GPU would consume 2GB of 6GB VRAM. Combined with other models, total could exceed 6GB. Draft02 estimates ~1.5GB per frame for XFeat/SuperPoint + LightGlue, but doesn't account for DINOv2 or faiss.
-
-### Budget Analysis
-Concurrent peak VRAM (worst case):
- XFeat inference: ~200MB
- LightGlue ONNX (FP16): ~500MB
- DINOv2 ViT-B/14: ~600MB
- SuperPoint: ~400MB
- faiss GPU: ~2GB
- ONNX Runtime overhead: ~300MB
- **Total: ~4.0GB** (without faiss GPU: ~2.0GB)
-
-Sequential loading (recommended):
- Step 1: XFeat + XFeat matcher (VO): ~400MB
- Step 2: DINOv2-S (coarse retrieval): ~300MB → unload
- Step 3: SuperPoint + LightGlue ONNX (fine matching): ~900MB
- Peak: ~1.3GB (with model switching)
-
-### Conclusion
-Use DINOv2 ViT-S/14 (300MB, 0.05s/img — fast enough for coarse retrieval). Run faiss on CPU (the embedding vectors are small, CPU search is <1ms for ~2000 vectors). Sequential model loading for GPU: VO models first, then satellite matching models. Keep models loaded but process sequentially (no concurrent GPU inference).
-
-### Confidence
-✅ High — based on documented VRAM numbers
-
---
-
-## Dimension 3: Rotation Handling Completeness
-
-### Fact Confirmation
-According to Fact #4, SuperPoint+LightGlue fail at 90°/180° rotations. According to Fact #6 (ISPRS 2025), SIFT+LightGlue outperforms SuperPoint+LightGlue in high-rotation UAV scenarios.
-
-### Problem
-Draft02 says "estimate heading from VO chain, rectify images before satellite matching, SIFT fallback for rotation-heavy cases." But:
-1. At segment start, there's no heading from VO chain — no rectification possible.
-2. No criteria for when to trigger SIFT fallback.
-3. Multi-rotation matching strategy ({0°, 90°, 180°, 270°}) not mentioned.
-
-### Conclusion
-Three-tier rotation handling:
-1. **Segment start (no heading)**: Try DINOv2 coarse retrieval (more rotation-robust than local features) → if match found, estimate heading from satellite alignment → proceed normally.
-2. **Normal operation (heading available)**: Rectify to approximate north-up using accumulated heading → SuperPoint+LightGlue.
-3. **Match failure fallback**: Try 4 rotations {0°, 90°, 180°, 270°} with SuperPoint. If still fails → SIFT+LightGlue (rotation-invariant).
-
-Trigger for SIFT: SuperPoint inlier ratio < 0.15 after rotation retry.
-
-### Confidence
-✅ High — based on confirmed LightGlue limitation + proven SIFT+LightGlue alternative
-
---
-
-## Dimension 4: VO Robustness (Homography Decomposition)
-
-### Fact Confirmation
-According to Fact #13, cv2.decomposeHomographyMat returns 4 solutions. Can return non-orthogonal matrices. Calibration matrix K precision is critical.
-
-### Problem
-Draft02 specifies selection by "motion consistent with previous direction + positive depth." This is underspecified for the first frame pair in a segment (no previous direction). Non-orthogonal R detection is missing.
-
-### Conclusion
-Disambiguation procedure:
-1. Compute all 4 decompositions.
-2. **Filter by positive depth**: triangulate a few matched points, reject solutions where points are behind camera.
-3. **Filter by plane normal**: for downward-looking camera, the normal should approximately point up (positive z component in camera frame).
-4. **Motion consistency**: if previous direction available, prefer solution consistent with expected motion direction.
-5. **Orthogonality check**: verify R'R ≈ I, det(R) ≈ 1. If not, re-orthogonalize via SVD.
-6. For first frame pair: rely on filters 2+3 only.
-
-### Confidence
-✅ High — based on well-documented decomposition ambiguity
-
---
-
-## Dimension 5: Satellite Coarse Retrieval
-
-### Fact Confirmation
-According to Fact #15, DINOv2+SALAD outperforms raw DINOv2 CLS token for retrieval. According to Fact #5, DINOv2 achieves 86.27 R@1 with adaptive enhancement.
-
-### Problem
-Draft02 proposes "DINOv2 global retrieval + faiss cosine similarity." Using raw CLS token is suboptimal. SALAD or patch-level feature aggregation would improve retrieval accuracy.
-
-### Conclusion
-DINOv2+SALAD is the better approach but adds a training/fine-tuning dependency. For a production system without the ability to fine-tune: use DINOv2 patch tokens (not just CLS) with spatial pooling, then cosine similarity via faiss. This captures more spatial information than CLS alone. If time permits, train SALAD head (30 minutes on appropriate dataset).
-
-Alternatively, consider SatDINO (DINOv2 pre-trained on satellite imagery) if available as a checkpoint.
-
-### Confidence
-⚠️ Medium — SALAD is proven but adding training dependency may not be worth the complexity for this use case
-
---
-
-## Dimension 6: Concurrency & Pipeline Architecture
-
-### Fact Confirmation
-Single GPU (RTX 2060) cannot run two models concurrently. Fact #8 shows sequential model inference times. Fact #3 shows LightGlue ONNX at ~50-100ms.
-
-### Problem
-Draft02 says satellite matching is "async — don't block VO pipeline" but on a single GPU, you can't parallelize GPU inference.
-
-### Conclusion
-Pipeline design:
-1. **VO (synchronous, per-frame)**: XFeat extract + match (~30ms total) → homography estimation (~5ms) → GTSAM update (~5ms) → emit position via SSE. **Total: ~40ms per frame.**
-2. **Satellite matching (asynchronous, overlapped with next frame's VO)**: DINOv2 coarse (~50ms) → SuperPoint+LightGlue fine (~150ms) → GTSAM update (~5ms) → emit refined position. **Total: ~205ms but overlapped.**
-3. **I/O (fully async)**: Tile download, DEM fetch, cache management — all via asyncio.
-4. **CPU tasks (parallel)**: faiss search (CPU), homography RANSAC (CPU-bound but fast).
-
-The GPU processes frames sequentially. The "async" part is that satellite matching for frame N happens while VO for frame N+1 proceeds. Since satellite matching (~205ms) is longer than VO (~40ms), the pipeline is satellite-matching-bound but VO results stream immediately.
-
-### Confidence
-✅ High — based on documented inference times
-
---
-
-## Dimension 7: Security Attack Surface
-
-### Fact Confirmation
-According to Fact #14, Pillow CVE-2025-48379 affects image loading. Fact #12 confirms SSE cleanup issues.
-
-### Problem
-Draft02 has JWT + rate limiting + CORS but misses:
- Image format/magic byte validation before loading
- Pillow version pinning
- Memory-limited image loading (a 100,000 × 100,000 pixel image could OOM)
- SSE heartbeat for connection health
- No mention of directory traversal prevention depth
-
-### Conclusion
-Additional security measures:
-1. Pin Pillow ≥11.3.0 in requirements.
-2. Validate image magic bytes (JPEG/PNG/TIFF) before loading with PIL.
-3. Check image dimensions before loading: reject if either dimension > 10,000px.
-4. Use OpenCV for loading (separate from PIL) — validate separately.
-5. Resolve image_folder path to canonical form (os.path.realpath) and verify it's under allowed base directories.
-6. Add Content-Security-Policy headers.
-7. SSE heartbeat every 15s to detect stale connections.
-8. Implement asyncio.Queue-based event publisher for SSE.
-
-### Confidence
-✅ High — based on documented CVE + known SSE issues
-
---
-
-## Dimension 8: Drift Management Strategy
-
-### Fact Confirmation
-According to Fact #17, NaviLoc demonstrates that trajectory-level optimization with noisy VPR measurements achieves 16x better accuracy than per-frame approaches. SatLoc-Fusion uses adaptive confidence metrics.
-
-### Problem
-Draft02's "drift limit factor" as a GTSAM custom factor is problematic: (1) custom Python factors are slow, (2) if no satellite anchors arrive for extended period, the drift factor has nothing to constrain against.
-
-### Conclusion
-Replace GTSAM drift factor with Segment Manager logic:
-1. Track cumulative VO displacement since last satellite anchor.
-2. If cumulative displacement > 100m without anchor: emit warning SSE event, increase satellite matching frequency/radius.
-3. If cumulative displacement > 200m: request user input with timeout.
-4. If cumulative displacement > 500m: mark segment as LOW confidence, continue but warn.
-5. Confidence score per position: decays exponentially with distance from nearest anchor.
-
-This is simpler, faster, and more controllable than a GTSAM custom factor.
-
-### Confidence
-✅ High — engineering judgment supported by SatLoc-Fusion's confidence-based approach
@@ -1,100 +0,0 @@
-# Validation Log — Draft02 Assessment (Draft03)
-
-## Validation Scenario 1: Factor graph initialization with first satellite match
-
-**Scenario**: Flight starts, VO processes 10 frames, satellite match arrives for frame 5.
-
-**Expected with Draft03 fixes**:
-1. GTSAM graph starts with PriorFactorPose2 at starting GPS (frame 0).
-2. BetweenFactorPose2 added for frames 0→1, 1→2, ..., 9→10.
-3. Satellite match for frame 5: add PriorFactorPose2 with position from satellite match and noise proportional to reprojection error × GSD.
-4. iSAM2.update() triggers backward correction — frames 0-4 and 5-10 both adjust.
-5. All positions in local ENU coordinates, converted to WGS84 for output.
-
-**Validation result**: Consistent. PriorFactorPose2 correctly constrains Pose2 variables. No GPSFactor API mismatch.
-
-## Validation Scenario 2: Segment start with unknown heading (rotation handling)
-
-**Scenario**: After a sharp turn, new segment starts. First image has unknown heading.
-
-**Expected with Draft03 fixes**:
-1. VO triple check fails → segment break.
-2. New segment starts. No heading available.
-3. For satellite coarse retrieval: DINOv2-S processes unrotated image → top-5 tiles.
-4. For fine matching: try SuperPoint+LightGlue at 4 rotations {0°, 90°, 180°, 270°}.
-5. If match found: heading estimated from satellite alignment. Subsequent images rectified.
-6. If no match: try SIFT+LightGlue (rotation-invariant).
-7. If still no match: request user input.
-
-**Validation result**: Consistent. Three-tier fallback addresses the heading bootstrap problem.
-
-## Validation Scenario 3: VRAM budget during satellite matching
-
-**Scenario**: Processing frame with concurrent VO + satellite matching on RTX 2060 (6GB).
-
-**Expected with Draft03 fixes**:
-1. XFeat features already extracted for VO: ~200MB VRAM.
-2. DINOv2 ViT-S/14 loaded for coarse retrieval: ~300MB.
-3. After coarse retrieval, DINOv2 can be unloaded or kept resident.
-4. SuperPoint loaded for fine matching: ~400MB.
-5. LightGlue ONNX loaded: ~500MB.
-6. Peak if all loaded: ~1.4GB.
-7. ONNX Runtime workspace: ~300MB.
-8. Total peak: ~1.7GB — well within 6GB.
-9. faiss runs on CPU — no VRAM impact.
-
-**Validation result**: Consistent. VRAM budget is comfortable even without model unloading.
-
-## Validation Scenario 4: Extended satellite failure (50+ frames)
-
-**Scenario**: Flying over area with outdated/changed satellite imagery. Satellite matching fails for 80 consecutive frames (~8km).
-
-**Expected with Draft03 fixes**:
-1. Frames 1-10: normal VO, satellite matching fails. Cumulative drift increases.
-2. Frame ~10 (1km drift): warning SSE event emitted. Satellite search radius expanded.
-3. Frame ~20 (2km drift): user_input_needed SSE event. If user provides GPS → anchor + backward correction.
-4. If user doesn't respond within timeout: continue with LOW confidence.
-5. Frame ~50 (5km drift): positions marked as very low confidence.
-6. Confidence score per position decays exponentially from last anchor.
-7. If satellite finally matches at frame 80: massive backward correction. Refined events emitted.
-
-**Validation result**: Consistent. Explicit drift thresholds are more predictable than the custom GTSAM factor approach.
-
-## Validation Scenario 5: Google Maps session token management
-
-**Scenario**: Processing a 3000-image flight. Need ~2000 satellite tiles.
-
-**Expected with Draft03 fixes**:
-1. On job start: create Google Maps session with POST /v1/createSession (returns session token).
-2. Use session token in all tile requests for this session.
-3. Daily limit: 15,000 tiles → sufficient for single flight.
-4. Monthly limit: 100,000 → ~50 flights.
-5. At 80% daily limit (12,000): switch to Mapbox.
-6. Mapbox: 200,000/month → additional ~100 flights capacity.
-
-**Validation result**: Consistent. Session management addressed correctly.
-
-## Review Checklist
- [x] Draft conclusions consistent with fact cards
- [x] No important dimensions missed
- [x] No over-extrapolation
- [x] Conclusions actionable/verifiable
- [x] GPSFactor API mismatch verified against GTSAM docs
- [x] VRAM budget calculated with specific model variants
- [x] Rotation handling addresses segment start edge case
- [x] Drift management has concrete thresholds
-
-## Counterexamples
- **Very low-texture terrain** (uniform sand/snow): XFeat VO might fail even on consecutive frames. Mitigation: track texture score per image, warn when low.
- **Satellite imagery completely missing for region**: Both Google and Mapbox might have no data. Mitigation: user-provided tiles are highest priority.
- **Multiple concurrent GPU processes**: Another process using the GPU could reduce available VRAM. Mitigation: document exclusive GPU access requirement.
-
-## Conclusions Requiring No Revision
-All conclusions validated. Key improvements are well-supported:
-1. Correct GTSAM factor types (PriorFactorPose2 instead of GPSFactor)
-2. DINOv2 ViT-S/14 for VRAM efficiency
-3. Three-tier rotation handling
-4. Explicit drift thresholds in Segment Manager
-5. asyncio.Queue-based SSE publisher
-6. CPU-based faiss
-7. Session token management for Google Maps
@@ -1,74 +0,0 @@
-# Acceptance Criteria Assessment
-
-## Acceptance Criteria
-
-| Criterion | Our Values | Researched Values | Cost/Timeline Impact | Status |
-|-----------|-----------|-------------------|---------------------|--------|
-| Position accuracy (80% of photos) | ≤50m error | 15-150m achievable depending on method. SatLoc (2025): <15m with adaptive fusion. Mateos-Ramirez (2024): 142m mean at 1000m+ altitude. At 400m altitude with better GSD (~6cm/px) and satellite correction, ≤50m for 80% is realistic | Moderate — requires high-quality satellite imagery and robust feature matching pipeline | **Modified** — see notes on satellite imagery quality dependency |
-| Position accuracy (60% of photos) | ≤20m error | Achievable only with satellite-anchored corrections, not with VO alone. SatLoc reports <15m with satellite anchoring + VO fusion. Requires 0.3-0.5 m/px satellite imagery and good terrain texture | High — requires premium satellite imagery, robust cross-view matching, and careful calibration | **Modified** — add dependency on satellite correction frequency |
-| Outlier tolerance | 350m displacement between consecutive photos | At 400m altitude, image footprint is ~375x250m. A 350m displacement means near-zero overlap. VO will fail; system must rely on IMU dead-reckoning or satellite re-localization | Low — standard outlier detection can handle this | Modified — specify fallback strategy (IMU dead-reckoning + satellite re-matching) |
-| Sharp turn handling (partial overlap) | <200m drift, <70° angle, <5% overlap | Standard VO fails below ~20-30% overlap. With <5% overlap, feature matching between consecutive frames is unreliable. Requires satellite-based re-localization or IMU bridging | High — requires separate re-localization module | Modified — clarify: "70%" should likely be "70 degrees"; add IMU-bridge requirement |
-| Disconnected route segments | System should reconnect disconnected chunks | This is essentially a place recognition / re-localization problem. Solvable via satellite image matching for each new segment independently | High — core architectural requirement affecting system design | Modified — add: each segment should independently localize via satellite matching |
-| User fallback input | Ask user after 3 consecutive failures | Reasonable fallback. Needs UI/API integration for interactive input | Low | No change |
-| Processing time per image | <5 seconds | On Jetson Orin Nano Super (8GB shared memory): feasible with optimized pipeline. CUDA feature extraction ~50ms, matching ~100-500ms, satellite crop+match ~1-3s. Full pipeline 2-4s is achievable with image downsampling and TensorRT optimization | Moderate — requires TensorRT optimization and image downsampling strategy | **Modified** — specify this is for Jetson Orin Nano Super, not RTX 2060 |
-| Real-time streaming | SSE for immediate results + refinement | Standard pattern, well-supported | Low | No change |
-| Image Registration Rate | >95% | For consecutive frames with nadir camera in good conditions: 90-98% achievable. Drops significantly during sharp turns and over low-texture terrain (water, uniform fields). The 95% target conflicts with sharp-turn handling requirement | Moderate — requires learning-based matchers (SuperPoint/LightGlue) | **Modified** — clarify: 95% applies to "normal flight" segments only; sharp-turn frames are expected failures handled by re-localization |
-| Mean Reprojection Error | <1.0 pixels | Achievable with modern methods (LightGlue, SuperGlue). Traditional methods typically 1-3 px. Deep learning matchers routinely achieve 0.3-0.8 px with proper calibration | Moderate — requires deep learning feature matchers | No change — achievable |
-| REST API + SSE architecture | Background service | Standard architecture, well-supported in Python (FastAPI + SSE) | Low | No change |
-| Satellite imagery resolution | ≥0.5 m/px, ideally 0.3 m/px | Google Maps for eastern Ukraine: variable, typically 0.5-1.0 m/px in rural areas. 0.3 m/px unlikely from Google Maps. Commercial providers (Maxar, Planet) offer 0.3-0.5 m/px but at significant cost | **High** — Google Maps may not meet 0.5 m/px in all areas of the operational region. 0.3 m/px requires commercial satellite providers | **Modified** — current Google Maps limitation may make this unachievable for all areas; consider fallback for degraded satellite quality |
-| Confidence scoring | Per-position estimate (high=satellite, low=VO) | Standard practice in sensor fusion. Easy to implement | Low | No change |
-| Output format | WGS84, GeoJSON or CSV | Standard, trivial to implement | Negligible | No change |
-| Satellite imagery age | <2 years where possible | Google Maps imagery for conflict zones (eastern Ukraine) may be significantly outdated or intentionally degraded. Recency is hard to guarantee | Medium — may need multiple satellite sources | **Modified** — flag: conflict zone imagery may be intentionally limited |
-| Max VO cumulative drift | <100m between satellite corrections | VIO drift typically 0.8-1% of distance. Between corrections at 1km intervals: ~10m drift. 100m budget allows corrections every ~10km — very generous | Low — easily achievable if corrections happen at reasonable intervals | No change — generous threshold |
-| Memory usage | <8GB shared memory (Jetson Orin Nano Super) | Binding constraint. 8GB LPDDR5 shared between CPU and GPU. ~6-7GB usable after OS. 26MP images need downsampling | **Critical** — all processing must fit within 8GB shared memory | **Updated** — changed to Jetson Orin Nano Super constraint |
-| Object center coordinates | Accuracy consistent with frame-center accuracy | New criterion — derives from problem statement requirement | Low — once frame position is known, object position follows from pixel offset + GSD | **Added** |
-| Sharp turn handling | <200m drift, <70 degrees, <5% overlap. 95% registration rate applies to normal flight only | Clarified from original "70%" to "70 degrees". Split registration rate expectation | Low — clarification only | **Updated** |
-| Offline preprocessing time | Not time-critical (minutes/hours before flight) | New criterion — no constraint existed | Low | **Added** |
-
-## Restrictions Assessment
-
-| Restriction | Our Values | Researched Values | Cost/Timeline Impact | Status |
-|-------------|-----------|-------------------|---------------------|--------|
-| Aircraft type | Fixed-wing only | Appropriate — fixed-wing has predictable motion model, mostly forward flight. Simplifies VO assumptions | N/A | No change |
-| Camera mount | Downward-pointing, fixed, not autostabilized | Implies roll/pitch affect image. At 400m altitude, moderate roll/pitch causes manageable image shift. IMU data can compensate. Non-stabilized means more variable image overlap and orientation | Medium — must use IMU data for image dewarping or accept orientation-dependent accuracy | **Modified** — add: IMU-based image orientation correction should be considered |
-| Operational region | Eastern/southern Ukraine (left of Dnipro) | Conflict zone — satellite imagery may be degraded, outdated, or restricted. Terrain: mix of agricultural, urban, forest. Agricultural areas have seasonal texture changes | **High** — satellite imagery availability and quality is a significant risk | **Modified** — flag operational risk: imagery access in conflict zones |
-| Image resolution | FullHD to 6252x4168, known camera parameters | 26MP at max is large for edge processing. Must downsample for feature extraction. Known camera intrinsics enable proper projective geometry | Medium — pipeline must handle variable resolutions | No change |
-| Altitude | Predefined, ≤1km, terrain height negligible | At 400m: GSD ~6cm/px, footprint ~375x250m. Terrain "negligible" is an approximation — even 50m terrain variation at 400m altitude causes ~12% scale error. The referenced paper (Mateos-Ramirez 2024) shows terrain elevation is a primary error source | **Medium** — "terrain height negligible" needs qualification. At 400m, terrain variations >50m become significant | **Modified** — add: terrain height can be neglected only if variations <50m within image footprint |
-| IMU data availability | "A lot of data from IMU" | IMU provides: accelerometer, gyroscope, magnetometer. Crucial for: dead-reckoning during feature-less frames, image orientation compensation, scale estimation, motion prediction. Standard tactical IMUs provide 100-400Hz data | Low — standard IMU integration | **Modified** — specify: IMU data includes gyroscope + accelerometer at ≥100Hz; will be used for orientation compensation and dead-reckoning fallback |
-| Weather | Mostly sunny | Favorable for visual methods. Shadows can actually help feature matching. Reduces image quality variability | Low — favorable condition | No change |
-| Satellite provider | Google Maps (potentially outdated) | **Critical limitation**: Google Maps satellite API has usage limits, unknown update frequency for eastern Ukraine, potential conflict-zone restrictions. Resolution may not meet 0.5 m/px in rural areas. No guarantee of recency | **High** — single-provider dependency is a significant risk | **Modified** — consider: (1) downloading tiles ahead of time for the operational area, (2) having a fallback provider strategy |
-| Photo count | Up to 3000, typically 500-1500 | At 3fps and 500-1500 photos: 3-8 minutes of flight. At ~100m spacing: 50-150km route. Memory for 3000 pre-extracted satellite feature maps needs careful management on 8GB | Medium — batch processing and memory management needed | **Modified** — add: pipeline must manage memory for up to 3000 frames on 8GB device |
-| Sharp turns | Next photo may have no common objects with previous | This is the hardest edge case. Complete visual discontinuity requires satellite-based re-localization. IMU provides heading/velocity for bridging. System must be architected around this possibility | High — drives core architecture decision | No change — already captured as a defining constraint |
-| Processing hardware | Jetson Orin Nano Super, 67 TOPS | 8GB shared LPDDR5, 1024 CUDA cores, 32 Tensor Cores, 102 GB/s bandwidth. TensorRT for inference optimization. Power: 7-25W. Significantly less capable than desktop GPU | **Critical** — all processing must fit within 8GB shared memory, pipeline must be optimized for TensorRT | **Modified** — CONTRADICTS AC's RTX 2060 reference. Must be the binding constraint |
-
-## Key Findings
-
-1. **CRITICAL CONTRADICTION**: The AC mentions "RTX 2060 compatibility" (16GB RAM + 6GB VRAM) but the restriction specifies Jetson Orin Nano Super (8GB shared memory). These are fundamentally different platforms. **The Jetson must be the binding constraint.** All processing, including model weights, image buffers, and intermediate results, must fit within ~6-7GB usable memory (OS takes ~1-1.5GB).
-
-2. **Satellite Imagery Risk**: Google Maps as the sole satellite provider for a conflict zone in eastern Ukraine presents significant quality, resolution, and recency risks. The 0.3 m/px "ideal" resolution is unlikely available from Google Maps for this region. The system design must be robust to degraded satellite reference quality (0.5-1.0 m/px).
-
-3. **Accuracy is Achievable but Conditional**: The 50m/80% and 20m/60% accuracy targets are achievable based on recent research (SatLoc 2025: <15m with adaptive fusion), but **only when satellite corrections are successful**. VO-only segments will drift ~1% of distance traveled. The system must maximize satellite correction frequency.
-
-4. **Sharp Turn Handling Drives Architecture**: The requirement to handle disconnected route segments with no visual overlap between consecutive frames means the system cannot rely solely on sequential VO. It must have an independent satellite-based geo-localization capability for each frame or segment — this is a core architectural requirement.
-
-5. **Processing Time is Feasible**: <5s per image on Jetson Orin Nano Super is achievable with: (a) image downsampling (e.g., to 2000x1300), (b) TensorRT-optimized models, (c) efficient satellite region cropping. GPU-accelerated feature extraction takes ~50ms, matching ~100-500ms, satellite matching ~1-3s.
-
-6. **Missing AC: Object Center Coordinates**: The problem statement mentions "coordinates of the center of any object in these photos" but no acceptance criterion specifies the accuracy requirement for this. Need to add.
-
-7. **Missing AC: DEM/Elevation Data**: Research shows terrain elevation is a primary error source for pixel-to-meter conversion at these altitudes. If terrain variations are >50m, a DEM is needed. No current restriction mentions DEM availability.
-
-8. **Missing AC: Offline Preprocessing Time**: No constraint on how long satellite image preprocessing can take before the flight.
-
-9. **"70%" in Sharp Turn AC is Ambiguous**: "at an angle of less than 70%" — this likely means 70 degrees, not 70%.
-
-## Sources
-
- SatLoc: Hierarchical Adaptive Fusion Framework for GNSS-denied UAV Localization (2025) — <15m error, >90% coverage, 2+ Hz on edge hardware
- Mateos-Ramirez et al. "Visual Odometry in GPS-Denied Zones for Fixed-Wing UAV" (2024) — 142.88m mean error over 17km at 1000m+ altitude, 0.83% error rate with satellite correction
- NVIDIA Jetson Orin Nano Super specs: 8GB LPDDR5, 67 TOPS, 1024 CUDA cores, 102 GB/s bandwidth
- cuda-efficient-features: Feature extraction benchmarks — 4K in ~12ms on Jetson Xavier
- SIFT+LightGlue for UAV image mosaicking (ISPRS 2025) — superior performance across diverse scenarios
- SuperPoint+LightGlue comparative analysis (2024) — best balance of robustness, accuracy, efficiency
- Google Maps satellite resolution: 0.15m-30m depending on location and source imagery
- VIO drift benchmarks: 0.82-1% of distance traveled (EuRoC, outdoor flights)
- UAVSAR cross-modality matching: 1.83-2.86m RMSE with deep learning approach (Springer 2026)
@@ -1,88 +0,0 @@
-# Question Decomposition
-
-## Original Question
-Research the GPS-denied onboard navigation problem for a fixed-wing UAV and find the best solution architecture. The system must determine frame-center GPS coordinates using visual odometry, satellite image matching, and IMU fusion — all running on a Jetson Orin Nano Super (8GB shared memory, 67 TOPS).
-
-## Active Mode
-Mode A Phase 2 — Initial Research (Problem & Solution)
-
-## Rationale
-No existing solution drafts. Full problem decomposition and solution research needed.
-
-## Problem Context Summary (from INPUT_DIR)
- **Platform**: Fixed-wing UAV, camera pointing down (not stabilized), 400m altitude max 1km
- **Camera**: ADTi Surveyor Lite 26S v2, 26MP (6252x4168), focal length 25mm, sensor width 23.5mm
- **GSD at 400m**: ~6cm/pixel, footprint ~375x250m
- **Frame rate**: 3 fps (interval ~333ms, real-world could be 400-500ms)
- **Photo count**: 500-3000 per flight
- **IMU**: Available at high rate
- **Initial GPS**: Known; GPS may be denied/spoofed during flight
- **Satellite reference**: Pre-uploaded Google Maps tiles
- **Hardware**: Jetson Orin Nano Super, 8GB shared memory, 67 TOPS
- **Region**: Eastern/southern Ukraine (conflict zone)
- **Key challenge**: Reconnecting disconnected route segments after sharp turns
-
-## Question Type Classification
-**Decision Support** — we need to evaluate and select the best architectural approach and component technologies for each part of the pipeline.
-
-## Research Subject Boundary Definition
-
-| Dimension | Boundary |
-|-----------|----------|
-| Population | Fixed-wing UAVs with nadir cameras at 200-1000m altitude |
-| Geography | Rural/semi-urban terrain in eastern Ukraine |
-| Timeframe | Current state-of-the-art (2023-2026) |
-| Level | Edge computing (Jetson-class, 8GB memory), real-time processing |
-
-## Decomposed Sub-Questions
-
-### A. Existing/Competitor Solutions
-1. What existing systems solve GPS-denied UAV visual navigation?
-2. What open-source implementations exist for VO + satellite matching?
-3. What commercial/military solutions address this problem?
-
-### B. Architecture Components
-4. What is the optimal pipeline architecture (sequential vs parallel, streaming)?
-5. How should VO, satellite matching, and IMU fusion be combined (loosely vs tightly coupled)?
-6. How to handle disconnected route segments (the core architectural challenge)?
-
-### C. Visual Odometry Component
-7. What VO algorithms work best for aerial nadir imagery on edge hardware?
-8. What feature extractors/matchers are optimal for Jetson (SuperPoint, ORB, XFeat)?
-9. How to handle scale estimation with known altitude and camera parameters?
-10. What is the optimal image downsampling strategy for 26MP on 8GB memory?
-
-### D. Satellite Image Matching Component
-11. How to efficiently match UAV frames against pre-loaded satellite tiles?
-12. What cross-view matching methods work for aerial-to-satellite registration?
-13. How to preprocess and index satellite tiles for fast retrieval?
-14. How to handle resolution mismatch (6cm UAV vs 50cm+ satellite)?
-
-### E. IMU Fusion Component
-15. How to fuse IMU data with visual estimates (EKF, UKF, factor graph)?
-16. How to use IMU for dead-reckoning during feature-less frames?
-17. How to use IMU for image orientation compensation (non-stabilized camera)?
-
-### F. Edge Optimization
-18. How to fit the full pipeline in 8GB shared memory?
-19. What TensorRT optimizations are available for feature extractors?
-20. How to achieve <5s per frame on Jetson Orin Nano Super?
-
-### G. API & Streaming
-21. What is the best approach for REST API + SSE on Python/Jetson?
-22. How to implement progressive result refinement?
-
-## Timeliness Sensitivity Assessment
-
- **Research Topic**: GPS-denied UAV visual navigation with edge processing
- **Sensitivity Level**: 🟠 High
- **Rationale**: Deep learning feature matchers (SuperPoint, LightGlue, XFeat) and edge inference frameworks (TensorRT) evolve rapidly. Jetson Orin Nano Super is a recent (Dec 2024) product. Cross-view geo-localization is an active research area.
- **Source Time Window**: 12 months (prioritize 2025-2026)
- **Priority official sources to consult**:
-  1. NVIDIA Jetson documentation and benchmarks
-  2. OpenCV / kornia / hloc official docs
-  3. Recent papers on cross-view geo-localization (CVPR, ECCV, ICCV 2024-2025)
- **Key version information to verify**:
-  - JetPack SDK: Current version ____
-  - SuperPoint/LightGlue: Latest available for TensorRT ____
-  - XFeat: Version and Jetson compatibility ____
@@ -1,151 +0,0 @@
-# Source Registry
-
-## Source #1
- **Title**: Visual Odometry in GPS-Denied Zones for Fixed-Wing UAV with Reduced Accumulative Error Based on Satellite Imagery
- **Link**: https://www.mdpi.com/2076-3417/14/16/7420
- **Tier**: L1
- **Publication Date**: 2024-08-22
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Fixed-wing UAV GPS-denied navigation
- **Research Boundary Match**: ✅ Full match
- **Summary**: VO + satellite correction pipeline for fixed-wing UAV at 1000m+ altitude. Mean error 142.88m over 17km (0.83%). Uses ORB features, centroid-based displacement, Kalman filter smoothing, quadtree for satellite keypoint indexing.
- **Related Sub-question**: A1, B5, C7, D11
-
-## Source #2
- **Title**: SatLoc: Hierarchical Adaptive Fusion Framework for GNSS-denied UAV Localization
- **Link**: https://www.scilit.com/publications/e5cafaf875a49297a62b298a89d5572f
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV localization in GNSS-denied environments
- **Research Boundary Match**: ✅ Full match
- **Summary**: Three-layer fusion: DinoV2 for satellite geo-localization, XFeat for VO, optical flow for velocity. Adaptive confidence-based weighting. <15m error, >90% coverage, 2+ Hz on edge hardware.
- **Related Sub-question**: B4, B5, C8, D12
-
-## Source #3
- **Title**: XFeat: Accelerated Features for Lightweight Image Matching (CVPR 2024)
- **Link**: https://arxiv.org/abs/2404.19174
- **Tier**: L1
- **Publication Date**: 2024-04
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Edge device feature matching
- **Research Boundary Match**: ✅ Full match
- **Summary**: 5x faster than SuperPoint, runs on CPU at VGA resolution. Sparse and semi-dense matching. TensorRT deployment available for Jetson. Comparable accuracy to SuperPoint.
- **Related Sub-question**: C8, F18, F20
-
-## Source #4
- **Title**: XFeat TensorRT Implementation
- **Link**: https://github.com/PranavNedunghat/XFeatTensorRT
- **Tier**: L2
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Summary**: C++ TensorRT implementation of XFeat, tested on Jetson Orin NX 16GB with JetPack 6.0, CUDA 12.2, TensorRT 8.6.
- **Related Sub-question**: C8, F18, F19
-
-## Source #5
- **Title**: SuperPoint+LightGlue TensorRT Deployment
- **Link**: https://github.com/fettahyildizz/superpoint_lightglue_tensorrt
- **Tier**: L2
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Summary**: C++ TensorRT implementation of SuperPoint+LightGlue. Production-ready deployment for Jetson platforms.
- **Related Sub-question**: C8, F19
-
-## Source #6
- **Title**: FP8 Quantized LightGlue in TensorRT
- **Link**: https://fabio-sim.github.io/blog/fp8-quantized-lightglue-tensorrt-nvidia-model-optimizer/
- **Tier**: L2
- **Publication Date**: 2026
- **Timeliness Status**: ✅ Currently valid
- **Summary**: Up to ~6x speedup with FP8 quantization. Requires Hopper/Ada Lovelace GPUs (not available on Jetson Orin Nano Ampere). FP16 is the best available precision for Orin Nano.
- **Related Sub-question**: F19
-
-## Source #7
- **Title**: NVIDIA JetPack 6.2 Release Notes
- **Link**: https://docs.nvidia.com/jetson/archives/jetpack-archived/jetpack-62/release-notes/index.html
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: CUDA 12.6.10, TensorRT 10.3.0, cuDNN 9.3. Super Mode for Orin Nano: up to 2x inference performance, 50% memory bandwidth boost. Power modes: 15W, 25W, MAXN SUPER.
- **Related Sub-question**: F18, F19, F20
-
-## Source #8
- **Title**: cuda-efficient-features (GPU feature detection benchmarks)
- **Link**: https://github.com/fixstars/cuda-efficient-features
- **Tier**: L2
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Summary**: 4K detection: 12ms on Jetson Xavier. 8K: 27.5ms. 40K keypoints extraction: 20-25ms on Xavier. Orin Nano Super should be comparable or better.
- **Related Sub-question**: F20
-
-## Source #9
- **Title**: Adaptive Covariance Hybrid EKF/UKF for Visual-Inertial Odometry
- **Link**: https://arxiv.org/abs/2512.17505
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: Hybrid EKF/UKF achieves 49% better position accuracy, 57% better rotation accuracy than ESKF alone, at 48% lower computational cost than full UKF. Includes adaptive sensor confidence scoring.
- **Related Sub-question**: E15
-
-## Source #10
- **Title**: SIFT+LightGlue for UAV Image Mosaicking (ISPRS 2025)
- **Link**: https://isprs-archives.copernicus.org/articles/XLVIII-2-W11-2025/169/2025/
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: SIFT+LightGlue outperforms SuperPoint+LightGlue for UAV mosaicking across diverse scenarios. Superior in both low-texture and high-texture environments.
- **Related Sub-question**: C8, D12
-
-## Source #11
- **Title**: UAVision - GNSS-Denied UAV Visual Localization System
- **Link**: https://github.com/ArboriseRS/UAVision
- **Tier**: L4
- **Publication Date**: 2024-2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: Open-source system using LightGlue for map matching. Includes image processing modules and visualization.
- **Related Sub-question**: A2
-
-## Source #12
- **Title**: TerboucheHacene/visual_localization
- **Link**: https://github.com/TerboucheHacene/visual_localization
- **Tier**: L4
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Summary**: Vision-based GNSS-free localization with SuperPoint/SuperGlue/GIM matching. Optimized VO + satellite image matching hybrid pipeline. Learning-based matchers for natural environments.
- **Related Sub-question**: A2, D12
-
-## Source #13
- **Title**: GNSS-Denied Geolocalization with Terrain Constraints
- **Link**: https://github.com/yfs90/gnss-denied-uav-geolocalization
- **Tier**: L4
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Summary**: No altimeters/IMU required, uses image matching + terrain constraints. GPS-comparable accuracy for day/night across varied terrain.
- **Related Sub-question**: A2
-
-## Source #14
- **Title**: Google Maps Tile API Documentation
- **Link**: https://developers.google.com/maps/documentation/tile/satellite
- **Tier**: L1
- **Publication Date**: Current
- **Timeliness Status**: ✅ Currently valid
- **Summary**: Zoom levels 0-22. Satellite tiles via HTTPS. Session tokens required. Bulk download possible but subject to usage policies.
- **Related Sub-question**: D13
-
-## Source #15
- **Title**: NaviLoc: Trajectory-Level Visual Localization for GNSS-Denied UAV Navigation
- **Link**: https://www.mdpi.com/2504-446X/10/2/97
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Summary**: Trajectory-level optimization rather than per-frame matching. Optimizes entire trajectory against satellite reference for improved accuracy.
- **Related Sub-question**: B4, D11
-
-## Source #16
- **Title**: GSD Estimation for UAV Photogrammetry
- **Link**: https://blog.truegeometry.com/calculators/UAV_photogrammetry_workflows_calculation.html
- **Tier**: L3
- **Publication Date**: Current
- **Timeliness Status**: ✅ Currently valid
- **Summary**: GSD = (sensor_width × altitude) / (focal_length × image_width). For our case: (23.5mm × 400m) / (25mm × 6252) = 0.06 m/pixel.
- **Related Sub-question**: C9
@@ -1,121 +0,0 @@
-# Fact Cards
-
-## Fact #1
- **Statement**: XFeat achieves up to 5x faster inference than SuperPoint while maintaining comparable accuracy for pose estimation. It runs in real-time on CPU at VGA resolution.
- **Source**: Source #3 (CVPR 2024 paper)
- **Phase**: Phase 2
- **Target Audience**: Edge device deployments
- **Confidence**: ✅ High
- **Related Dimension**: Feature Extraction
-
-## Fact #2
- **Statement**: XFeat TensorRT implementation exists and is tested on Jetson Orin NX 16GB with JetPack 6.0, CUDA 12.2, TensorRT 8.6.
- **Source**: Source #4
- **Phase**: Phase 2
- **Target Audience**: Jetson platform deployment
- **Confidence**: ✅ High
- **Related Dimension**: Feature Extraction, Edge Optimization
-
-## Fact #3
- **Statement**: SatLoc framework achieves <15m absolute localization error with >90% trajectory coverage at 2+ Hz on edge hardware, using DinoV2 for satellite matching, XFeat for VO, and optical flow for velocity.
- **Source**: Source #2
- **Phase**: Phase 2
- **Target Audience**: GNSS-denied UAV localization
- **Confidence**: ⚠️ Medium (paper details not fully accessible)
- **Related Dimension**: Overall Architecture, Accuracy
-
-## Fact #4
- **Statement**: Mateos-Ramirez et al. achieved 142.88m mean error over 17km (0.83% error rate) with VO + satellite correction on a fixed-wing UAV at 1000m+ altitude. Without satellite correction, error accumulated to 850m+ over 17km.
- **Source**: Source #1
- **Phase**: Phase 2
- **Target Audience**: Fixed-wing UAV at high altitude
- **Confidence**: ✅ High
- **Related Dimension**: Accuracy, Architecture
-
-## Fact #5
- **Statement**: VIO systems typically drift 0.8-1% of distance traveled. Between satellite corrections at 1km intervals, expected drift is ~10m.
- **Source**: Multiple sources (arxiv VIO benchmarks)
- **Phase**: Phase 2
- **Target Audience**: Aerial VIO systems
- **Confidence**: ✅ High
- **Related Dimension**: VO Drift
-
-## Fact #6
- **Statement**: Jetson Orin Nano Super: 8GB LPDDR5 shared memory, 1024 CUDA cores, 32 Tensor Cores, 102 GB/s bandwidth, 67 TOPS INT8. JetPack 6.2: CUDA 12.6.10, TensorRT 10.3.0.
- **Source**: Source #7
- **Phase**: Phase 2
- **Target Audience**: Hardware specification
- **Confidence**: ✅ High
- **Related Dimension**: Edge Optimization
-
-## Fact #7
- **Statement**: CUDA-accelerated feature detection at 4K (3840x2160): ~12ms on Jetson Xavier. At 8K: ~27.5ms. Descriptor extraction for 40K keypoints: ~20-25ms on Xavier. Orin Nano Super has comparable or slightly better compute.
- **Source**: Source #8
- **Phase**: Phase 2
- **Target Audience**: Jetson GPU performance
- **Confidence**: ✅ High
- **Related Dimension**: Processing Time
-
-## Fact #8
- **Statement**: Hybrid EKF/UKF achieves 49% better position accuracy than ESKF alone at 48% lower computational cost than full UKF. Includes adaptive sensor confidence scoring based on image entropy and motion blur.
- **Source**: Source #9
- **Phase**: Phase 2
- **Target Audience**: VIO fusion
- **Confidence**: ✅ High
- **Related Dimension**: Sensor Fusion
-
-## Fact #9
- **Statement**: SIFT+LightGlue outperforms SuperPoint+LightGlue for UAV mosaicking across diverse scenarios (low-texture agricultural and high-texture urban).
- **Source**: Source #10
- **Phase**: Phase 2
- **Target Audience**: UAV image matching
- **Confidence**: ✅ High
- **Related Dimension**: Feature Matching
-
-## Fact #10
- **Statement**: GSD for our system at 400m: (23.5mm × 400m) / (25mm × 6252px) = 0.060 m/pixel. Image footprint: 6252 × 0.06 = 375m width, 4168 × 0.06 = 250m height.
- **Source**: Source #16 + camera parameters
- **Phase**: Phase 2
- **Target Audience**: Our specific system
- **Confidence**: ✅ High
- **Related Dimension**: Scale Estimation
-
-## Fact #11
- **Statement**: Google Maps satellite tiles available via Tile API at zoom levels 0-22. Max zoom varies by region. For eastern Ukraine, zoom 18 (~0.6 m/px) is typically available; zoom 19 (~0.3 m/px) may not be.
- **Source**: Source #14
- **Phase**: Phase 2
- **Target Audience**: Satellite imagery
- **Confidence**: ⚠️ Medium (exact zoom availability for eastern Ukraine unverified)
- **Related Dimension**: Satellite Reference
-
-## Fact #12
- **Statement**: FP8 quantization for LightGlue requires Hopper/Ada GPUs. Jetson Orin Nano uses Ampere architecture — limited to FP16 as best TensorRT precision.
- **Source**: Source #6, Source #7
- **Phase**: Phase 2
- **Target Audience**: Jetson optimization
- **Confidence**: ✅ High
- **Related Dimension**: Edge Optimization
-
-## Fact #13
- **Statement**: SuperPoint+LightGlue TensorRT C++ deployment is available and production-tested. ONNX Runtime path achieves 2-4x speedup over compiled PyTorch.
- **Source**: Source #5, Source #6
- **Phase**: Phase 2
- **Target Audience**: Production deployment
- **Confidence**: ✅ High
- **Related Dimension**: Feature Matching, Edge Optimization
-
-## Fact #14
- **Statement**: Cross-view matching (UAV-to-satellite) is fundamentally harder than same-view matching due to extreme viewpoint differences. Deep learning embeddings (DinoV2, CLIP-based) are the state-of-the-art for coarse retrieval. Local features are used for fine alignment.
- **Source**: Multiple (Sources #2, #12, #15)
- **Phase**: Phase 2
- **Target Audience**: Cross-view geo-localization
- **Confidence**: ✅ High
- **Related Dimension**: Satellite Matching
-
-## Fact #15
- **Statement**: Quadtree spatial indexing enables O(log n) nearest-neighbor lookup for satellite keypoints. Combined with GeoHash for fast region encoding, this is the standard approach for tile management.
- **Source**: Sources #1, #14
- **Phase**: Phase 2
- **Target Audience**: Spatial indexing
- **Confidence**: ✅ High
- **Related Dimension**: Satellite Tile Management
@@ -1,71 +0,0 @@
-# Comparison Framework
-
-## Selected Framework Type
-Decision Support — evaluating solution options per component
-
-## Architecture Components to Evaluate
-
-1. Feature Extraction & Matching (VO frame-to-frame)
-2. Satellite Image Matching (cross-view geo-registration)
-3. Sensor Fusion (VO + satellite + IMU)
-4. Satellite Tile Preprocessing & Indexing
-5. Image Downsampling Strategy
-6. Re-localization (disconnected segments)
-7. API & Streaming Layer
-
-## Component 1: Feature Extraction & Matching (VO)
-
-| Dimension | XFeat | SuperPoint + LightGlue | ORB (OpenCV) |
-|-----------|-------|----------------------|--------------|
-| Speed (Jetson) | ~2-5ms per frame (VGA), 5x faster than SuperPoint | ~15-50ms per frame (VGA, TensorRT FP16) | ~5-10ms per frame (CUDA) |
-| Accuracy | Comparable to SuperPoint on pose estimation | State-of-the-art for local features | Lower accuracy, not scale-invariant |
-| Memory | <100MB model | ~200-400MB model+inference | Negligible |
-| TensorRT support | Yes (C++ impl available for Jetson Orin NX) | Yes (C++ impl available) | N/A (native CUDA) |
-| Cross-view capability | Limited (same-view designed) | Better with LightGlue attention | Poor for cross-view |
-| Rotation invariance | Moderate | Good with LightGlue | Good (by design) |
-| Jetson validation | Tested on Orin NX (JetPack 6.0) | Tested on multiple Jetson platforms | Native OpenCV CUDA |
-| **Fit for VO** | ✅ Best — fast, accurate, Jetson-proven | ⚠️ Good but heavier | ⚠️ Fast but less accurate |
-| **Fit for satellite matching** | ⚠️ Moderate | ✅ Better for cross-view with attention | ❌ Poor for cross-view |
-
-## Component 2: Satellite Image Matching (Cross-View)
-
-| Dimension | Local Feature Matching (SIFT/SuperPoint + LightGlue) | Global Descriptor Retrieval (DinoV2/CLIP) | Template Matching (NCC) |
-|-----------|-----------------------------------------------------|------------------------------------------|------------------------|
-| Approach | Extract keypoints in both UAV and satellite images, match descriptors | Encode both images into global vectors, compare by distance | Slide UAV image over satellite tile, compute correlation |
-| Accuracy | Sub-pixel when matches found (best for fine alignment) | Tile-level (~50-200m depending on tile size) | Pixel-level but sensitive to appearance changes |
-| Speed | ~100-500ms for match+geometric verification | ~50-100ms for descriptor comparison | ~500ms-2s for large search area |
-| Robustness to viewpoint | Good with LightGlue attention | Excellent (trained for cross-view) | Poor (requires similar viewpoint) |
-| Memory | ~300-500MB (model + keypoints) | ~200-500MB (model) | Low |
-| Failure rate | High in low-texture, seasonal changes | Lower — semantic understanding | High in changed scenes |
-| **Recommended role** | Fine alignment (after coarse retrieval) | Coarse retrieval (select candidate tile) | Not recommended |
-
-## Component 3: Sensor Fusion
-
-| Dimension | EKF (Extended Kalman Filter) | Error-State EKF (ESKF) | Hybrid ESKF/UKF | Factor Graph (GTSAM) |
-|-----------|-------------------------------|------------------------|------------------|---------------------|
-| Accuracy | Baseline | Better for rotation | 49% better than ESKF | Best overall |
-| Compute cost | Lowest | Low | 48% less than full UKF | Highest |
-| Implementation complexity | Low | Medium | Medium-High | High |
-| Handles non-linearity | Linearization errors | Better for small errors | Best among KF variants | Full non-linear |
-| Real-time on Jetson | ✅ | ✅ | ✅ | ⚠️ Depends on graph size |
-| Multi-rate sensor support | Manual | Manual | Manual | Native |
-| **Fit** | ⚠️ Baseline option | ✅ Good starting point | ✅ Best KF option | ⚠️ Overkill for this system |
-
-## Component 4: Satellite Tile Management
-
-| Dimension | GeoHash + In-Memory | Quadtree + Memory-Mapped Files | Pre-extracted Feature DB |
-|-----------|--------------------|-----------------------------|------------------------|
-| Lookup speed | O(1) hash | O(log n) tree traversal | O(1) hash + feature load |
-| Memory usage | All tiles in RAM | On-demand loading | Features only (smaller) |
-| Preprocessing | Fast | Moderate | Slow (extract all features offline) |
-| Flexibility | Fixed grid | Adaptive resolution | Fixed per-tile |
-| **Fit for 8GB** | ❌ Too much RAM for large areas | ✅ Memory-efficient | ✅ Best — smallest footprint |
-
-## Component 5: Image Downsampling Strategy
-
-| Dimension | Fixed Resize (e.g., 1600x1066) | Pyramid (multi-scale) | ROI-based (center crop + full) |
-|-----------|-------------------------------|----------------------|-------------------------------|
-| Speed | Fast, single scale | Slower, multiple passes | Medium |
-| Accuracy | Good if GSD ratio maintained | Best for multi-scale features | Good for center, loses edges |
-| Memory | ~5MB per frame | ~7-8MB per frame | ~6MB per frame |
-| **Fit** | ✅ Best tradeoff | ⚠️ Unnecessary complexity | ⚠️ Loses coverage |
@@ -1,129 +0,0 @@
-# Reasoning Chain
-
-## Dimension 1: Feature Extraction for Visual Odometry
-
-### Fact Confirmation
-XFeat is 5x faster than SuperPoint (Fact #1), has TensorRT deployment on Jetson (Fact #2), and comparable accuracy for pose estimation. SatLoc (the most relevant state-of-the-art system) uses XFeat for its VO component (Fact #3).
-
-### Reference Comparison
-SuperPoint+LightGlue is more accurate for cross-view matching but heavier. ORB is fast but less accurate and not robust to appearance changes. SIFT+LightGlue is best for mosaicking (Fact #9) but slower.
-
-### Conclusion
-**XFeat for VO (frame-to-frame)** — it's the fastest learned feature, Jetson-proven, and used by the closest state-of-the-art system (SatLoc). For satellite matching, a different approach is needed because cross-view matching requires viewpoint-invariant features.
-
-### Confidence
-✅ High — supported by SatLoc architecture and CVPR 2024 benchmarks.
-
---
-
-## Dimension 2: Satellite Image Matching Strategy
-
-### Fact Confirmation
-Cross-view matching is fundamentally harder than same-view (Fact #14). Deep learning embeddings (DinoV2) are state-of-the-art for coarse retrieval (Fact #3). Local features are better for fine alignment. SatLoc uses DinoV2 for satellite matching specifically.
-
-### Reference Comparison
-A two-stage coarse-to-fine approach is the dominant pattern in literature: (1) global descriptor retrieves candidate region, (2) local feature matching refines position. Pure local-feature matching has high failure rate for cross-view due to extreme viewpoint differences.
-
-### Conclusion
-**Two-stage approach**: (1) Coarse — use a lightweight global descriptor to find the best-matching satellite tile within the search area (VO-predicted position ± uncertainty radius). (2) Fine — use local feature matching (SuperPoint+LightGlue or XFeat) between UAV frame and the matched satellite tile to get precise position. The coarse stage can also serve as the re-localization mechanism for disconnected segments.
-
-### Confidence
-✅ High — consensus across multiple recent papers and the SatLoc system.
-
---
-
-## Dimension 3: Sensor Fusion Approach
-
-### Fact Confirmation
-Hybrid ESKF/UKF achieves 49% better accuracy than ESKF alone at 48% lower cost than full UKF (Fact #8). Factor graphs (GTSAM) offer the best accuracy but are computationally expensive.
-
-### Reference Comparison
-For our system: IMU runs at 100-400Hz, VO at ~3Hz (frame rate), satellite corrections at variable rate (whenever matching succeeds). We need multi-rate fusion that handles intermittent satellite corrections and continuous IMU.
-
-### Conclusion
-**Error-State EKF (ESKF)** as the baseline fusion approach — it's well-understood, lightweight, handles multi-rate sensors naturally, and is proven for VIO on edge hardware. Upgrade to hybrid ESKF/UKF if ESKF accuracy is insufficient. Factor graphs are overkill for this real-time edge system.
-
-The filter state: position (lat/lon), velocity, orientation (quaternion), IMU biases. Measurements: VO-derived displacement (high rate), satellite-derived absolute position (variable rate), IMU (highest rate for prediction).
-
-### Confidence
-✅ High — ESKF is the standard choice for embedded VIO systems.
-
---
-
-## Dimension 4: Satellite Tile Preprocessing & Indexing
-
-### Fact Confirmation
-Quadtree enables O(log n) lookups (Fact #15). Pre-extracting features offline saves runtime compute. 8GB memory limits in-memory tile storage.
-
-### Reference Comparison
-Full tiles in memory is infeasible for large areas. Memory-mapped files allow on-demand loading. Pre-extracted feature databases have the smallest runtime footprint.
-
-### Conclusion
-**Offline preprocessing pipeline**:
-1. Download Google Maps satellite tiles at max zoom (18-19) for the operational area
-2. Extract features (XFeat or SuperPoint) from each tile
-3. Compute global descriptors (lightweight, e.g., NetVLAD or cosine-pooled XFeat descriptors) per tile
-4. Store: tile metadata (GPS bounds, zoom level), features + descriptors in a GeoHash-indexed database
-5. Build spatial index (GeoHash) for fast lookup by GPS region
-
-**Runtime**: Given VO-estimated position, query GeoHash to find nearby tiles, compare global descriptors for coarse match, then local feature matching for fine alignment.
-
-### Confidence
-✅ High — standard approach used by all relevant systems.
-
---
-
-## Dimension 5: Image Downsampling Strategy
-
-### Fact Confirmation
-26MP images need downsampling for 8GB device (Fact #6). Feature extraction at 4K takes ~12ms on Jetson Xavier (Fact #7). UAV GSD at 400m is ~6cm/px (Fact #10). Satellite GSD is ~60cm/px at zoom 18.
-
-### Reference Comparison
-For VO (frame-to-frame): features at full resolution are wasteful — consecutive frames at 6cm GSD overlap ~80%, and features at lower resolution are sufficient for displacement estimation. For satellite matching: we need to match at satellite resolution (~60cm/px), so downsampling to match satellite GSD is natural.
-
-### Conclusion
-**Downsample to ~1600x1066** (factor ~4x each dimension). This yields ~24cm/px GSD — still 2.5x finer than satellite, sufficient for feature matching. Image size: ~5MB (RGB). Feature extraction at this resolution: <10ms. This is the single resolution for both VO and satellite matching.
-
-### Confidence
-✅ High — standard practice for edge processing of high-res imagery.
-
---
-
-## Dimension 6: Disconnected Segment Handling
-
-### Fact Confirmation
-SatLoc uses satellite matching as an independent localization source that works regardless of VO state (Fact #3). The AC requires reconnecting disconnected segments as a core capability.
-
-### Reference Comparison
-Pure VO cannot handle zero-overlap transitions. IMU dead-reckoning bridges short gaps (seconds). Satellite-based re-localization provides absolute position regardless of VO state.
-
-### Conclusion
-**Independent satellite localization per frame** — every frame attempts satellite matching regardless of VO state. This naturally handles disconnected segments:
-1. When VO succeeds: satellite matching refines position (high confidence)
-2. When VO fails (sharp turn): satellite matching provides absolute position (sole source)
-3. When both fail: IMU dead-reckoning with low confidence score
-4. After 3 consecutive total failures: request user input
-
-Segment reconnection is automatic: all positions are in the same global (WGS84) frame via satellite matching. No explicit "reconnection" needed — segments share the satellite reference.
-
-### Confidence
-✅ High — this is the key architectural insight.
-
---
-
-## Dimension 7: Processing Pipeline Architecture
-
-### Fact Confirmation
-<5s per frame required (AC). Feature extraction ~10ms, VO matching ~20-50ms, satellite coarse retrieval ~50-100ms, satellite fine matching ~200-500ms, fusion ~1ms. Total: ~300-700ms per frame.
-
-### Conclusion
-**Pipelined parallel architecture**:
- Thread 1 (Camera): Capture frame, downsample, extract features → push to queue
- Thread 2 (VO): Match with previous frame, compute displacement → push to fusion
- Thread 3 (Satellite): Search nearby tiles, coarse retrieval, fine matching → push to fusion
- Thread 4 (Fusion): ESKF prediction (IMU), update (VO), update (satellite) → emit result via SSE
-
-VO and satellite matching can run in parallel for each frame. Fusion integrates results as they arrive. This enables <1s per frame total latency.
-
-### Confidence
-✅ High — standard producer-consumer pipeline.
@@ -1,98 +0,0 @@
-# Validation Log
-
-## Validation Scenario 1: Normal Flight (80% of time)
-UAV flies straight, consecutive frames overlap ~70-80%. Terrain has moderate texture (agricultural + urban mix).
-
-### Expected Based on Conclusions
- XFeat extracts features in ~5ms, VO matching in ~20ms
- Satellite matching succeeds: coarse retrieval ~50ms, fine matching ~300ms
- ESKF fuses both: position accuracy ~10-20m (satellite-anchored)
- Total processing: <500ms per frame
- Confidence: HIGH
-
-### Actual Validation (against literature)
-SatLoc reports <15m error with >90% coverage under similar conditions. Mateos-Ramirez reports 0.83% drift with satellite correction. Both align with our expected performance.
-
-### Result: ✅ PASS
-
---
-
-## Validation Scenario 2: Sharp Turn (5-10% of time)
-UAV makes a 60-degree turn. Next frame has <5% overlap with previous. Heading changes rapidly.
-
-### Expected Based on Conclusions
- VO fails (insufficient feature overlap) — detected by low match count
- IMU provides heading and approximate displacement for ~1-2 frames
- Satellite matching attempts independent localization of the new frame
- If satellite match succeeds: position recovered, segment continues
- If satellite match fails: IMU dead-reckoning with LOW confidence
-
-### Potential Issues
- Satellite matching may also fail if the frame is heavily tilted (non-nadir view during turn)
- IMU drift during turn: at 100m/s for 1s, displacement ~100m. IMU drift over 1s: ~1-5m — acceptable
-
-### Result: ⚠️ CONDITIONAL PASS — depends on satellite matching success during turn. Non-stabilized camera may produce tilted images that are harder to match. IMU provides reasonable bridge.
-
---
-
-## Validation Scenario 3: Disconnected Route (rare, <5%)
-UAV completes segment A, makes a 90+ degree turn, flies a new heading. Segment B has no overlap with segment A. Multiple such segments possible.
-
-### Expected Based on Conclusions
- Each segment independently localizes via satellite matching
- No explicit reconnection needed — all in WGS84 frame
- Per-segment accuracy depends on satellite matching success rate
- Low-confidence gaps between segments until satellite match succeeds
-
-### Result: ✅ PASS — architecture handles this natively via independent per-frame satellite matching.
-
---
-
-## Validation Scenario 4: Memory-Constrained Operation (always)
-3000 frames, 8GB shared memory. Full pipeline running.
-
-### Expected Based on Conclusions
- Downsampled frame: ~5MB per frame. Keep 2 in memory (current + previous): ~10MB
- XFeat model (TensorRT): ~50-100MB
- Satellite tile features (loaded tiles): ~200-500MB for tiles near current position
- ESKF state: <1MB
- OS + runtime: ~1.5GB
- Total: ~2-3GB active, well within 8GB
-
-### Potential Issues
- Satellite feature DB for large operational areas could be large on disk (not memory — loaded on demand)
- Need careful management of tile loading/unloading
-
-### Result: ✅ PASS — 8GB is sufficient with proper memory management.
-
---
-
-## Validation Scenario 5: Degraded Satellite Imagery
-Google Maps tiles at 0.5-1.0 m/px resolution. Some areas have outdated imagery. Seasonal appearance changes.
-
-### Expected Based on Conclusions
- Coarse retrieval (global descriptors) should handle moderate appearance changes
- Fine matching may fail on outdated/seasonal tiles — confidence drops to LOW
- System falls back to VO + IMU in degraded areas
- Multiple consecutive failures → user input request
-
-### Potential Issues
- If large areas have degraded satellite imagery, the system may operate mostly in VO+IMU mode with significant drift
- 50m accuracy target may not be achievable in these areas
-
-### Result: ⚠️ CONDITIONAL PASS — system degrades gracefully, but accuracy targets depend on satellite quality. This is a known risk per Phase 1 assessment.
-
---
-
-## Review Checklist
- [x] Draft conclusions consistent with fact cards
- [x] No important dimensions missed
- [x] No over-extrapolation
- [x] Conclusions actionable/verifiable
- [x] Sharp turn handling addressed
- [x] Memory constraints validated
- [ ] Issue: Satellite imagery quality in eastern Ukraine remains a risk
- [ ] Issue: Non-stabilized camera during turns may degrade satellite matching
-
-## Conclusions Requiring No Revision
-All major architectural decisions validated. Two known risks (satellite quality, non-stabilized camera during turns) are acknowledged and handled by the fallback hierarchy.
@@ -1,80 +0,0 @@
-# Question Decomposition — Solution Assessment (Mode B)
-
-## Original Question
-Assess the existing solution draft (solution_draft01.md) for weak points, security vulnerabilities, and performance bottlenecks, then produce a revised solution draft.
-
-## Active Mode
-Mode B: Solution Assessment — `solution_draft01.md` exists and is the highest-numbered draft.
-
-## Question Type Classification
- **Primary**: Problem Diagnosis — identify weak points, vulnerabilities, bottlenecks in existing solution
- **Secondary**: Decision Support — evaluate alternatives for identified issues
-
-## Research Subject Boundary Definition
-
-| Dimension | Boundary |
-|-----------|----------|
-| **Domain** | GPS-denied UAV visual navigation, aerial geo-referencing |
-| **Geography** | Eastern/southern Ukraine (left of Dnipro River) — steppe terrain, potential conflict-related satellite imagery degradation |
-| **Hardware** | Desktop/laptop with NVIDIA RTX 2060+, 16GB RAM, 6GB VRAM |
-| **Software** | Python ecosystem, GPU-accelerated CV/ML |
-| **Timeframe** | Current state-of-the-art (2024-2026), production-ready tools |
-| **Scale** | 500-3000 images per flight, up to 6252×4168 resolution |
-
-## Problem Context Summary
- UAV aerial photos taken consecutively ~100m apart, camera pointing down (not autostabilized)
- Only starting GPS known — must determine GPS for all subsequent images
- Must handle: sharp turns, outlier photos (up to 350m gap), disconnected route segments
- Processing <5s/image, real-time SSE streaming, REST API service
- No IMU data available
-
-## Decomposed Sub-Questions
-
-### A: Cross-View Matching Viability
-"Is SuperPoint+LightGlue with perspective warping reliable for UAV-to-satellite cross-view matching, or are there specialized cross-view methods that would perform better?"
-
-### B: Homography-Based VO Robustness
-"Is homography-based VO (flat terrain assumption) robust enough for non-stabilized camera with potential roll/pitch variations and non-flat objects?"
-
-### C: Satellite Imagery Reliability
-"What are the risks of relying solely on Google Maps satellite imagery for eastern Ukraine, and what fallback strategies exist?"
-
-### D: Processing Time Feasibility
-"Are the processing time estimates (<5s per image) realistic on RTX 2060 with SuperPoint+LightGlue+satellite matching pipeline?"
-
-### E: Optimizer Specification
-"Is the sliding window optimizer well-specified, and are there more proven alternatives like factor graph optimization?"
-
-### F: Camera Rotation Handling
-"How should the system handle arbitrary image rotation from non-stabilized camera mount?"
-
-### G: Security Assessment
-"What are the security vulnerabilities in the REST API + SSE architecture with image processing pipeline?"
-
-### H: Newer Tools & Libraries
-"Are there newer (2025-2026) tools, models, or approaches that outperform the current selections (SuperPoint, LightGlue, etc.)?"
-
-### I: Segment Management Robustness
-"Is the segment management strategy robust enough for multiple disconnected segments, especially when satellite anchoring fails for a segment?"
-
-### J: Memory & Resource Management
-"Can the pipeline stay within 16GB RAM / 6GB VRAM while processing 3000 images at 6252×4168 resolution?"
-
---
-
-## Timeliness Sensitivity Assessment
-
- **Research Topic**: GPS-denied UAV visual navigation using learned feature matching and satellite geo-referencing
- **Sensitivity Level**: 🟠 High
- **Rationale**: Computer vision feature matching models (SuperPoint, LightGlue, etc.) are actively evolving with new versions and competitors. However, the core algorithms (homography, VO, optimization) are stable. The tool ecosystem changes frequently.
- **Source Time Window**: 12 months (2025-2026)
- **Priority official sources to consult**:
-  1. LightGlue / SuperPoint GitHub repos (releases, issues)
-  2. OpenCV documentation (current version)
-  3. Google Maps Tiles API documentation
-  4. Recent aerial geo-referencing papers (2024-2026)
- **Key version information to verify**:
-  - LightGlue: current version and ONNX/TensorRT support status
-  - SuperPoint: current version and alternatives
-  - FastAPI: SSE support status
-  - Google Maps Tiles API: pricing, coverage, rate limits
@@ -1,201 +0,0 @@
-# Source Registry — Solution Assessment (Mode B)
-
-## Source #1
- **Title**: GLEAM: Learning to Match and Explain in Cross-View Geo-Localization
- **Link**: https://arxiv.org/abs/2509.07450
- **Tier**: L1
- **Publication Date**: 2025-09
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Cross-view geo-localization researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Framework for cross-view geo-localization with explainable matching across modalities. Demonstrates that specialized cross-view methods outperform generic feature matchers.
-
-## Source #2
- **Title**: Robust UAV Image Mosaicking Using SIFT and LightGlue (ISPRS 2025)
- **Link**: https://isprs-archives.copernicus.org/articles/XLVIII-2-W11-2025/169/2025/
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV photogrammetry and aerial image processing
- **Research Boundary Match**: ✅ Full match
- **Summary**: SIFT+LightGlue achieves superior spatial consistency and reliability for UAV image mosaicking, including low-texture and high-rotation conditions. SIFT outperforms SuperPoint for rotation-heavy scenarios.
-
-## Source #3
- **Title**: Precise GPS-Denied UAV Self-Positioning via Context-Enhanced Cross-View Geo-Localization (CEUSP)
- **Link**: https://arxiv.org/abs/2502.11408 / https://github.com/eksnew/ceusp
- **Tier**: L1
- **Publication Date**: 2025-02
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: GPS-denied UAV navigation
- **Research Boundary Match**: ⚠️ Partial overlap (urban, not steppe)
- **Summary**: DINOv2-based cross-view matching for UAV self-positioning. State-of-the-art on DenseUAV benchmark. Uses retrieval-based (not feature-matching) approach.
-
-## Source #4
- **Title**: SatLoc Dataset and Hierarchical Adaptive Fusion Framework
- **Link**: https://www.mdpi.com/2072-4292/17/17/3048
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: GNSS-denied UAV navigation
- **Research Boundary Match**: ✅ Full match
- **Summary**: Three-layer architecture: DINOv2 for absolute geo-localization, XFeat for VO, optical flow for velocity. Adaptive fusion with confidence weighting. <15m absolute error on edge hardware.
-
-## Source #5
- **Title**: LightGlue ONNX/TensorRT acceleration blog
- **Link**: https://fabio-sim.github.io/blog/accelerating-lightglue-inference-onnx-runtime-tensorrt/
- **Tier**: L2
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: LightGlue users optimizing inference
- **Research Boundary Match**: ✅ Full match
- **Summary**: LightGlue ONNX achieves 2-4x speedup over PyTorch. FP8 quantization (Ada/Hopper GPUs only) adds 6x more. RTX 2060 does NOT support FP8.
-
-## Source #6
- **Title**: LightGlue-ONNX GitHub repository
- **Link**: https://github.com/fabio-sim/LightGlue-ONNX
- **Tier**: L2
- **Publication Date**: 2024-2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: LightGlue deployment engineers
- **Research Boundary Match**: ✅ Full match
- **Summary**: ONNX export for LightGlue with FlashAttention-2 support. TopK-trick for ~30% speedup. Pre-exported models available.
-
-## Source #7
- **Title**: LightGlue GitHub Issue #64 — Rotation sensitivity
- **Link**: https://github.com/cvg/LightGlue/issues/64
- **Tier**: L4
- **Publication Date**: 2023-2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: LightGlue users
- **Research Boundary Match**: ✅ Full match
- **Summary**: LightGlue (with SuperPoint/DISK) is NOT rotation-invariant. 90° or 180° rotation causes matching failure. Manual rectification needed.
-
-## Source #8
- **Title**: LightGlue GitHub Issue #13 — No-match handling
- **Link**: https://github.com/cvg/LightGlue/issues/13
- **Tier**: L4
- **Publication Date**: 2023
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: LightGlue users
- **Research Boundary Match**: ✅ Full match
- **Summary**: LightGlue lacks explicit training on unmatchable pairs. May produce geometrically meaningless matches instead of rejecting non-overlapping views.
-
-## Source #9
- **Title**: YFS90/GNSS-Denied-UAV-Geolocalization GitHub
- **Link**: https://github.com/yfs90/gnss-denied-uav-geolocalization
- **Tier**: L1
- **Publication Date**: 2024-2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: GPS-denied UAV navigation
- **Research Boundary Match**: ✅ Full match
- **Summary**: <7m MAE using terrain-weighted constraint optimization + 2D-3D geo-registration. Uses DEM data. Validated across 20 complex scenarios. Works with publicly available satellite maps.
-
-## Source #10
- **Title**: Efficient image matching for UAV visual navigation via DALGlue (Scientific Reports 2025)
- **Link**: https://www.nature.com/articles/s41598-025-21602-5
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV visual navigation
- **Research Boundary Match**: ✅ Full match
- **Summary**: 11.8% MMA improvement over LightGlue. Uses dual-tree complex wavelet transform + adaptive spatial feature fusion + linear attention. Designed for UAV dynamic flight.
-
-## Source #11
- **Title**: XFeat: Accelerated Features for Lightweight Image Matching (CVPR 2024)
- **Link**: https://arxiv.org/html/2404.19174v1 / https://github.com/verlab/accelerated_features
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Real-time feature matching applications
- **Research Boundary Match**: ✅ Full match
- **Summary**: 5x faster than SuperPoint. Runs real-time on CPU. Sparse + semi-dense matching. Used by SatLoc-Fusion for VO. 1500+ GitHub stars.
-
-## Source #12
- **Title**: An Oblique-Robust Absolute Visual Localization Method (IEEE TGRS 2024)
- **Link**: https://ieeexplore.ieee.org/iel7/36/10354519/10356107.pdf
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: GPS-denied UAV localization
- **Research Boundary Match**: ✅ Full match
- **Summary**: SE(2)-steerable network for rotation-equivariant features. Handles drastic perspective changes, non-perpendicular camera angles. No additional training for new scenes.
-
-## Source #13
- **Title**: Google Maps Tiles API Usage and Billing
- **Link**: https://developers.google.com/maps/documentation/tile/usage-and-billing
- **Tier**: L1
- **Publication Date**: 2025-2026 (continuously updated)
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Google Maps API users
- **Research Boundary Match**: ✅ Full match
- **Summary**: 100,000 free tile requests/month. Rate limit: 6,000/min, 15,000/day for 2D tiles. $200/month free credit expired Feb 2025. Now pay-as-you-go only.
-
-## Source #14
- **Title**: GTSAM Python API and Factor Graph examples
- **Link**: https://github.com/borglab/gtsam / https://pypi.org/project/gtsam-develop/
- **Tier**: L1
- **Publication Date**: 2025-2026 (v4.2 stable, v4.3a1 dev)
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Robot navigation, SLAM
- **Research Boundary Match**: ✅ Full match
- **Summary**: Python bindings for factor graph optimization. GPSFactor for absolute position constraints. iSAM2 for incremental optimization. Stable v4.2 for production use.
-
-## Source #15
- **Title**: Copernicus DEM documentation
- **Link**: https://documentation.dataspace.copernicus.eu/APIs/SentinelHub/Data/DEM.html
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: DEM data users
- **Research Boundary Match**: ✅ Full match
- **Summary**: Free 30m DEM (GLO-30) covering Ukraine. API access via Sentinel Hub Process API. Registration required.
-
-## Source #16
- **Title**: Homography Decomposition Revisited (IJCV 2025)
- **Link**: https://link.springer.com/article/10.1007/s11263-025-02680-4
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Computer vision researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Existing homography decomposition methods can be unstable in certain configurations. Proposes hybrid framework for improved stability.
-
-## Source #17
- **Title**: Sliding window factor graph optimization for visual/inertial navigation (Cambridge 2020)
- **Link**: https://www.cambridge.org/core/services/aop-cambridge-core/content/view/523C7C41D18A8D7C159C59235DF502D0/
- **Tier**: L1
- **Publication Date**: 2020
- **Timeliness Status**: ✅ Currently valid (foundational method)
- **Target Audience**: Navigation system designers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Sliding-window factor graph optimization combines accuracy of graph optimization with efficiency of windowed approach. Superior to separate filtering or full batch optimization.
-
-## Source #18
- **Title**: SuperPoint feature extraction and matching benchmarks
- **Link**: https://preview-www.nature.com/articles/s41598-024-59626-y/tables/3
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Feature matching benchmarking
- **Research Boundary Match**: ✅ Full match
- **Summary**: SuperPoint+LightGlue: ~0.36±0.06s per image pair for extraction+matching on GPU. Competitive accuracy for satellite stereo scenarios.
-
-## Source #19
- **Title**: DINOv2-Based UAV Visual Self-Localization in Low-Altitude Urban Environments
- **Link**: https://ui.adsabs.harvard.edu/abs/2025IRAL...10.2080Y/
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV visual localization researchers
- **Research Boundary Match**: ⚠️ Partial overlap (urban, not steppe)
- **Summary**: DINOv2-based method achieves 86.27 R@1 on DenseUAV benchmark for cross-view matching. Integrates global-local feature enhancement.
-
-## Source #20
- **Title**: Mapbox Satellite Tiles and Pricing
- **Link**: https://docs.mapbox.com/data/tilesets/reference/mapbox-satellite/ / https://mapbox.com/pricing
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Map tile consumers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Mapbox offers satellite tiles up to 0.3m resolution (zoom 16+). 200,000 free vector tile requests/month. Unlimited offline downloads on pay-as-you-go. Multi-provider imagery (Maxar, Landsat, Sentinel).
@@ -1,161 +0,0 @@
-# Fact Cards — Solution Assessment (Mode B)
-
-## Fact #1
- **Statement**: LightGlue (with SuperPoint/DISK descriptors) is NOT rotation-invariant. Image pairs with 90° or 180° rotation produce very few or zero matches. Manual image rectification is required before matching.
- **Source**: Source #7 (LightGlue GitHub Issue #64)
- **Phase**: Assessment
- **Target Audience**: UAV systems with non-stabilized cameras
- **Confidence**: ✅ High (confirmed by LightGlue maintainers)
- **Related Dimension**: Cross-view matching robustness, camera rotation handling
-
-## Fact #2
- **Statement**: LightGlue lacks explicit training on unmatchable image pairs. When given non-overlapping views (e.g., after sharp turn), it may return semantically correct but geometrically meaningless matches instead of correctly rejecting the pair.
- **Source**: Source #8 (LightGlue GitHub Issue #13)
- **Phase**: Assessment
- **Target Audience**: Systems requiring segment detection (VO failure detection)
- **Confidence**: ✅ High (confirmed by LightGlue maintainers)
- **Related Dimension**: Segment management, VO failure detection
-
-## Fact #3
- **Statement**: SatLoc-Fusion achieves <15m absolute localization error using a three-layer hierarchical approach: DINOv2 for coarse absolute geo-localization, XFeat for high-frequency VO, optical flow for velocity estimation. Runs real-time on 6 TFLOPS edge hardware.
- **Source**: Source #4 (SatLoc-Fusion, Remote Sensing 2025)
- **Phase**: Assessment
- **Target Audience**: GPS-denied UAV systems
- **Confidence**: ✅ High (peer-reviewed, with dataset)
- **Related Dimension**: Architecture, localization accuracy, hierarchical matching
-
-## Fact #4
- **Statement**: XFeat is 5x faster than SuperPoint with comparable accuracy. Runs real-time on CPU. Supports both sparse and semi-dense matching. 1500+ GitHub stars, actively maintained.
- **Source**: Source #11 (CVPR 2024)
- **Phase**: Assessment
- **Target Audience**: Real-time feature extraction
- **Confidence**: ✅ High (peer-reviewed, CVPR 2024)
- **Related Dimension**: Processing speed, feature extraction
-
-## Fact #5
- **Statement**: SIFT+LightGlue achieves superior spatial consistency and reliability for UAV image mosaicking, including in low-texture and high-rotation conditions. SIFT is rotation-invariant unlike SuperPoint.
- **Source**: Source #2 (ISPRS 2025)
- **Phase**: Assessment
- **Target Audience**: UAV image matching
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: Feature extraction, rotation handling
-
-## Fact #6
- **Statement**: SuperPoint+LightGlue extraction+matching takes ~0.36±0.06s per image pair on GPU (unspecified GPU model). This is for standard resolution images, not 6000+ pixel width.
- **Source**: Source #18
- **Phase**: Assessment
- **Target Audience**: Performance planning
- **Confidence**: ⚠️ Medium (GPU model not specified, may not be RTX 2060)
- **Related Dimension**: Processing time
-
-## Fact #7
- **Statement**: LightGlue ONNX/TensorRT achieves 2-4x speedup over compiled PyTorch. FP8 quantization adds 6x more but requires Ada Lovelace or newer GPUs. RTX 2060 (Turing) does NOT support FP8 — limited to FP16/INT8 acceleration.
- **Source**: Source #5, #6 (LightGlue-ONNX blog and repo)
- **Phase**: Assessment
- **Target Audience**: RTX 2060 deployment
- **Confidence**: ✅ High (benchmarked by repo maintainer)
- **Related Dimension**: Processing time, hardware constraints
-
-## Fact #8
- **Statement**: YFS90 achieves <7m MAE using terrain-weighted constraint optimization + 2D-3D geo-registration with DEM data. Validated across 20 complex scenarios including plains, hilly terrain, urban/rural. Works with publicly available satellite maps and DEM data. Re-localization capability after failures.
- **Source**: Source #9 (YFS90 GitHub)
- **Phase**: Assessment
- **Target Audience**: GPS-denied UAV navigation
- **Confidence**: ✅ High (peer-reviewed, open source, 69★)
- **Related Dimension**: Optimization approach, DEM integration, accuracy
-
-## Fact #9
- **Statement**: Google Maps $200/month free credit expired February 28, 2025. Current free tier is 100,000 tile requests/month. Rate limits: 6,000 requests/min, 15,000 requests/day for 2D tiles.
- **Source**: Source #13 (Google Maps official docs)
- **Phase**: Assessment
- **Target Audience**: Cost planning
- **Confidence**: ✅ High (official documentation)
- **Related Dimension**: Cost, satellite imagery access
-
-## Fact #10
- **Statement**: Google Maps satellite imagery for eastern Ukraine is likely updated only every 3-5+ years due to: conflict zone (lower priority), geopolitical challenges, limited user demand. This may not meet the AC requirement of "less than 2 years old."
- **Source**: Multiple web sources on Google Maps update frequency
- **Phase**: Assessment
- **Target Audience**: Satellite imagery reliability
- **Confidence**: ⚠️ Medium (general guidelines, not Ukraine-specific confirmation)
- **Related Dimension**: Satellite imagery reliability
-
-## Fact #11
- **Statement**: Mapbox Satellite offers imagery up to 0.3m resolution at zoom 16+, sourced from Maxar, Landsat, Sentinel. 200,000 free vector tile requests/month. Unlimited offline downloads on pay-as-you-go. Potentially more diverse and recent imagery for Ukraine than Google Maps alone.
- **Source**: Source #20 (Mapbox docs)
- **Phase**: Assessment
- **Target Audience**: Alternative satellite providers
- **Confidence**: ✅ High (official documentation)
- **Related Dimension**: Satellite imagery reliability, cost
-
-## Fact #12
- **Statement**: Copernicus DEM GLO-30 provides free 30m resolution global elevation data including Ukraine. Accessible via Sentinel Hub API. Can be used for terrain-weighted optimization like YFS90.
- **Source**: Source #15 (Copernicus docs)
- **Phase**: Assessment
- **Target Audience**: DEM integration
- **Confidence**: ✅ High (official documentation)
- **Related Dimension**: Position optimizer, terrain constraints
-
-## Fact #13
- **Statement**: GTSAM v4.2 (stable) provides Python bindings with GPSFactor for absolute position constraints and iSAM2 for incremental optimization. Can model VO constraints, satellite anchor constraints, and drift limits in a unified factor graph.
- **Source**: Source #14 (GTSAM docs)
- **Phase**: Assessment
- **Target Audience**: Optimizer design
- **Confidence**: ✅ High (widely used in robotics)
- **Related Dimension**: Position optimizer
-
-## Fact #14
- **Statement**: DALGlue achieves 11.8% MMA improvement over LightGlue on MegaDepth benchmark. Specifically designed for UAV visual navigation with wavelet transform preprocessing for handling dynamic flight blur.
- **Source**: Source #10 (Scientific Reports 2025)
- **Phase**: Assessment
- **Target Audience**: Feature matching selection
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: Feature matching
-
-## Fact #15
- **Statement**: The oblique-robust AVL method (IEEE TGRS 2024) uses SE(2)-steerable networks for rotation-equivariant features. Handles drastic perspective changes and non-perpendicular camera angles for UAV-to-satellite matching. No retraining needed for new scenes.
- **Source**: Source #12 (IEEE TGRS 2024)
- **Phase**: Assessment
- **Target Audience**: Cross-view matching
- **Confidence**: ✅ High (peer-reviewed, IEEE)
- **Related Dimension**: Cross-view matching, rotation handling
-
-## Fact #16
- **Statement**: Homography decomposition can be unstable in certain configurations (2025 IJCV study). Non-planar objects (buildings, trees) violate planar assumption. For aerial images, dominant ground plane exists but RANSAC inlier ratio drops with non-planar content.
- **Source**: Source #16 (IJCV 2025)
- **Phase**: Assessment
- **Target Audience**: VO design
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: VO robustness
-
-## Fact #17
- **Statement**: Sliding-window factor graph optimization combines the accuracy of full graph optimization with the efficiency of windowed processing. Superior to either pure filtering or full batch optimization for real-time navigation.
- **Source**: Source #17 (Cambridge 2020)
- **Phase**: Assessment
- **Target Audience**: Optimizer design
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: Position optimizer
-
-## Fact #18
- **Statement**: SuperPoint is a fully-convolutional model — GPU memory scales linearly with image resolution. 6252×4168 input would require significant VRAM. Standard practice is to downscale to 1024-2048 long edge for feature extraction.
- **Source**: Source #18, SuperPoint docs
- **Phase**: Assessment
- **Target Audience**: Memory management
- **Confidence**: ✅ High (architectural fact)
- **Related Dimension**: Memory management, processing pipeline
-
-## Fact #19
- **Statement**: For GPS-denied UAV localization, hierarchical coarse-to-fine approaches (image retrieval → local feature matching) are state-of-the-art. Direct local feature matching alone fails when the search area is too large or viewpoint difference is too high.
- **Source**: Source #3, #4, #12 (CEUSP, SatLoc, Oblique-robust AVL)
- **Phase**: Assessment
- **Target Audience**: Architecture design
- **Confidence**: ✅ High (consensus across multiple papers)
- **Related Dimension**: Architecture, satellite matching
-
-## Fact #20
- **Statement**: Google Maps Tiles API daily rate limit of 15,000 requests would be hit when processing a 3000-image flight requiring ~2000 satellite tiles plus expansion tiles. Need to either pre-cache or use the per-minute limit (6,000/min) strategically across multiple days.
- **Source**: Source #13 (Google Maps docs)
- **Phase**: Assessment
- **Target Audience**: System design
- **Confidence**: ✅ High (official rate limits)
- **Related Dimension**: Satellite tile management, rate limiting
@@ -1,79 +0,0 @@
-# Comparison Framework — Solution Assessment (Mode B)
-
-## Selected Framework Type
-Problem Diagnosis + Decision Support
-
-## Identified Weak Points and Assessment Dimensions
-
-### Dimension 1: Cross-View Matching Strategy (UAV→Satellite)
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| Strategy | Direct SuperPoint+LightGlue matching with perspective warping | No coarse localization stage. Fails when VO drift is large. LightGlue not rotation-invariant. | Hierarchical: DINOv2/global retrieval → SuperPoint+LightGlue refinement | Fact #1, #2, #15, #19 |
-| Rotation handling | Not addressed | Non-stabilized camera = rotated images. SuperPoint/LightGlue fail at 90°/180° | Image rectification via VO-estimated heading, or rotation-invariant features (SIFT for fallback) | Fact #1, #5 |
-| Domain gap | Perspective warping only | Insufficient for seasonal/illumination/resolution differences | Multi-scale matching, DINOv2 for semantic retrieval, warping + matched features | Fact #3, #15 |
-
-### Dimension 2: Feature Extraction & Matching
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| VO features | SuperPoint (~80ms) | Adequate but not optimized for speed | XFeat (5x faster, CPU-capable) for VO; keep SuperPoint for satellite matching | Fact #4 |
-| Matching | LightGlue | Good baseline. DALGlue 11.8% better MMA. | LightGlue with ONNX optimization as primary. DALGlue for evaluation. | Fact #7, #14 |
-| Non-match detection | Not addressed | LightGlue returns false matches on non-overlapping pairs | Inlier ratio + match count threshold + geometric consistency check | Fact #2 |
-
-### Dimension 3: Visual Odometry Robustness
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| Geometric model | Homography (planar assumption) | Unstable for non-planar objects. Decomposition instability in certain configs. | Homography with RANSAC + high inlier ratio requirement. Essential matrix as fallback. | Fact #16 |
-| Scale estimation | GSD from altitude | Valid if altitude is constant. Terrain elevation changes not accounted for. | Integrate Copernicus DEM for terrain-corrected GSD | Fact #12 |
-| Camera rotation | Not addressed | Non-stabilized camera introduces roll/pitch | Estimate rotation from VO, apply rectification before satellite matching | Fact #1, #5 |
-
-### Dimension 4: Position Optimizer
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| Algorithm | scipy.optimize sliding window | Generic optimizer, no proper uncertainty modeling, no factor types | GTSAM factor graph with iSAM2 incremental optimization | Fact #13, #17 |
-| Terrain constraints | Not used | YFS90 achieves <7m with terrain weighting | Integrate DEM-based terrain constraints via Copernicus DEM | Fact #8, #12 |
-| Drift modeling | Max 100m between anchors | Single hard constraint, no probabilistic modeling | Per-VO-step uncertainty based on inlier ratio, propagated through factor graph | Fact #17 |
-
-### Dimension 5: Satellite Imagery Reliability
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| Provider | Google Maps only | Eastern Ukraine: 3-5 year update cycle. $200 credit expired. 15K/day rate limit. | Multi-provider: Google Maps primary + Mapbox fallback + pre-cached tiles | Fact #9, #10, #11, #20 |
-| Freshness | Assumed adequate | May not meet AC "< 2 years old" for conflict zone | Provider selection per-area. User can provide custom imagery. | Fact #10 |
-| Rate limiting | Not addressed | 15,000/day cap could block large flights | Progressive download with request budgeting. Pre-cache for known areas. | Fact #20 |
-
-### Dimension 6: Processing Time Budget
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| Target | <5s (claim <2s) | Per-frame pipeline: VO match + satellite match + optimization. Total could exceed budget. | XFeat for VO (~20ms). LightGlue ONNX for satellite (~100ms). Async satellite matching. | Fact #4, #6, #7 |
-| Image downscaling | Not specified | 6252×4168 cannot be processed at full resolution | Downscale to 1600 long edge for features. Keep full resolution for GSD calculation. | Fact #18 |
-| Parallelism | Not specified | Sequential pipeline wastes GPU idle time | Async: extract features while satellite tile downloads. Pipeline overlap. | — |
-
-### Dimension 7: Memory Management
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| Image loading | Not specified | 6252×4168 × 3ch = 78MB per raw image. 3000 images = 234GB. | Stream images one at a time. Keep only current + previous features in memory. | Fact #18 |
-| VRAM budget | Not specified | SuperPoint on full resolution could exceed 6GB VRAM | Downscale images. Batch size 1. Clear GPU cache between frames. | Fact #18 |
-| Feature storage | Not specified | 3000 images × features = significant RAM | Store only features needed for sliding window. Disk-backed for older frames. | — |
-
-### Dimension 8: Security
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| Authentication | API key mentioned | No implementation details. API key in query params = insecure. | JWT tokens for session auth. Short-lived tokens for SSE connections. | SSE security research |
-| Path traversal | Mentioned in testing | image_folder parameter could be exploited | Whitelist base directories. Validate path doesn't escape allowed root. | — |
-| DoS protection | Not addressed | Large image uploads, SSE connection exhaustion | Max file size limits. Connection pool limits. Request rate limiting. | — |
-| API key storage | env var mentioned | Adequate baseline | .env file + secrets manager in production. Never log API keys. | — |
-
-### Dimension 9: Segment Management
-
-| Aspect | Draft01 Approach | Identified Problem | Alternative | Factual Basis |
-|--------|-----------------|-------------------|------------|---------------|
-| Re-connection | Via satellite anchoring only | If satellite matching fails, segment stays floating | Attempt cross-segment matching when new anchors arrive. DEM-based constraint stitching. | Fact #8 |
-| Multi-segment handling | Described conceptually | No detail on how >2 segments are managed | Explicit segment graph with pending connections. Priority queue for unresolved segments. | — |
-| User input fallback | POST /jobs/{id}/anchor | Good design. Needs timeout/escalation for when user doesn't respond. | Add configurable timeout before continuing with VO-only estimate. | — |
@@ -1,145 +0,0 @@
-# Reasoning Chain — Solution Assessment (Mode B)
-
-## Dimension 1: Cross-View Matching Strategy
-
-### Fact Confirmation
-According to Fact #1, LightGlue is not rotation-invariant and fails on rotated images. According to Fact #2, it returns false matches on non-overlapping pairs. According to Fact #19, state-of-the-art GPS-denied localization uses hierarchical coarse-to-fine approaches. SatLoc-Fusion (Fact #3) achieves <15m with DINOv2 + XFeat + optical flow.
-
-### Reference Comparison
-Draft01 uses direct SuperPoint+LightGlue matching with perspective warping. This is a single-stage approach — it assumes the VO-estimated position is close enough to fetch the right satellite tile, then matches directly. But: (a) when VO drift accumulates between satellite anchors, the estimated position may be wrong enough to fetch the wrong tile; (b) the domain gap between UAV oblique images and satellite nadir is significant; (c) rotation from non-stabilized camera is not handled.
-
-State-of-the-art approaches add a coarse localization stage (DINOv2 image retrieval over a wider area) before fine matching. This makes satellite matching robust to larger VO drift.
-
-### Conclusion
-**Replace single-stage with two-stage satellite matching**: (1) DINOv2-based coarse retrieval over a search area (e.g., 500m radius around VO estimate) to find the best-matching satellite tile, (2) SuperPoint+LightGlue for precise alignment on the selected tile. Add image rotation normalization before matching. This is the most critical improvement.
-
-### Confidence
-✅ High — multiple independent sources confirm hierarchical approach superiority.
-
---
-
-## Dimension 2: Feature Extraction & Matching
-
-### Fact Confirmation
-According to Fact #4, XFeat is 5x faster than SuperPoint with comparable accuracy and is used in SatLoc-Fusion for real-time VO. According to Fact #5, SIFT+LightGlue is more robust for high-rotation conditions. According to Fact #14, DALGlue improves LightGlue MMA by 11.8% for UAV scenarios.
-
-### Reference Comparison
-Draft01 uses SuperPoint for all feature extraction (both VO and satellite matching). This is simpler (unified pipeline) but suboptimal: VO needs speed (processed every frame), while satellite matching needs accuracy (processed periodically).
-
-### Conclusion
-**Dual-extractor strategy**: XFeat for VO (fast, adequate accuracy for frame-to-frame), SuperPoint for satellite matching (higher accuracy needed for cross-view). LightGlue with ONNX/TensorRT optimization as matcher. SIFT as fallback for rotation-heavy scenarios. DALGlue is promising but too new for production — monitor.
-
-### Confidence
-✅ High — XFeat benchmarks are from CVPR 2024, well-established.
-
---
-
-## Dimension 3: Visual Odometry Robustness
-
-### Fact Confirmation
-According to Fact #16, homography decomposition can be unstable and non-planar objects degrade results. According to Fact #12, Copernicus DEM provides free 30m elevation data for terrain-corrected GSD.
-
-### Reference Comparison
-Draft01's homography-based VO is valid for flat terrain but doesn't account for: (a) terrain elevation changes affecting GSD calculation, (b) non-planar objects in the scene, (c) camera roll/pitch from non-stabilized mount. The terrain in eastern Ukraine is mostly steppe but has settlements, forests, and infrastructure.
-
-### Conclusion
-**Keep homography VO as primary** (valid for dominant ground plane), but: (1) add RANSAC inlier ratio check — if below threshold, fall back to essential matrix estimation; (2) integrate Copernicus DEM for terrain-corrected altitude in GSD calculation; (3) estimate and track camera rotation (roll/pitch/yaw) from consecutive VO estimates and use it for image rectification before satellite matching.
-
-### Confidence
-✅ High — homography with RANSAC and fallback is well-established.
-
---
-
-## Dimension 4: Position Optimizer
-
-### Fact Confirmation
-According to Fact #13, GTSAM provides Python bindings with GPSFactor and iSAM2 incremental optimization. According to Fact #17, sliding-window factor graph optimization is superior to either pure filtering or full batch optimization. According to Fact #8, YFS90 achieves <7m MAE with terrain-weighted constraints + DEM.
-
-### Reference Comparison
-Draft01 proposes scipy.optimize with a custom sliding window. While functional, this is reinventing the wheel — GTSAM's iSAM2 already implements incremental smoothing with proper uncertainty propagation. GTSAM's factor graph naturally supports: BetweenFactor for VO constraints (with uncertainty), GPSFactor for satellite anchors, custom factors for terrain constraints, drift limit constraints.
-
-### Conclusion
-**Replace scipy.optimize with GTSAM iSAM2 factor graph**. Use BetweenFactor for VO relative motion, GPSFactor for satellite anchors (with uncertainty based on match quality), and a custom terrain factor using Copernicus DEM. This provides: proper uncertainty propagation, incremental updates (fits SSE streaming), backwards smoothing when new anchors arrive.
-
-### Confidence
-✅ High — GTSAM is production-proven, stable v4.2 available via pip.
-
---
-
-## Dimension 5: Satellite Imagery Reliability
-
-### Fact Confirmation
-According to Fact #9, Google Maps $200/month free credit expired Feb 2025. Current free tier is 100K tiles/month. According to Fact #10, eastern Ukraine imagery may be 3-5+ years old. According to Fact #20, 15,000/day rate limit could be hit on large flights. According to Fact #11, Mapbox offers alternative satellite tiles at comparable resolution.
-
-### Reference Comparison
-Draft01 relies solely on Google Maps. Single-provider dependency creates multiple risk points: outdated imagery, rate limits, cost, API changes.
-
-### Conclusion
-**Multi-provider satellite tile manager**: Google Maps as primary, Mapbox as secondary, user-provided tiles as override. Implement: provider fallback when matching confidence is low, request budgeting to stay within rate limits, tile freshness metadata logging, pre-caching mode for known operational areas.
-
-### Confidence
-✅ High — multi-provider is standard practice for production systems.
-
---
-
-## Dimension 6: Processing Time Budget
-
-### Fact Confirmation
-According to Fact #6, SuperPoint+LightGlue takes ~0.36s per pair on GPU. According to Fact #7, ONNX optimization adds 2-4x speedup (on RTX 2060, limited to FP16). According to Fact #4, XFeat is 5x faster than SuperPoint for VO.
-
-### Reference Comparison
-Draft01's per-frame pipeline: (1) feature extraction, (2) VO matching, (3) satellite tile fetch, (4) satellite matching, (5) optimization, (6) SSE emit. Total estimated without optimization: ~1-2s for VO + ~0.5-1s for satellite + overhead = 2-4s. With ONNX optimization for matching and XFeat for VO, this drops to ~0.5-1.5s.
-
-### Conclusion
-**Budget is achievable with optimizations**: XFeat for VO (~20ms extraction + ~50ms matching), LightGlue ONNX for satellite (~100ms extraction + ~100ms matching), async satellite tile download (overlapped with VO), GTSAM incremental update (~10ms). Total: ~0.5-1s per frame. Satellite matching can be async — not every frame needs satellite match. Image downscaling to 1600 long edge is essential.
-
-### Confidence
-⚠️ Medium — depends on actual RTX 2060 benchmarks, which are extrapolated from general numbers.
-
---
-
-## Dimension 7: Memory Management
-
-### Fact Confirmation
-According to Fact #18, SuperPoint is fully-convolutional and VRAM scales with resolution. 6252×4168 images would require significant VRAM and RAM.
-
-### Reference Comparison
-Draft01 doesn't specify memory management. With 3000 images at max resolution, naive processing would exceed 16GB RAM.
-
-### Conclusion
-**Strict memory management**: (1) Downscale all images to max 1600 long edge before feature extraction; (2) stream images one at a time — only keep current + previous frame features in GPU memory; (3) store features for sliding window in CPU RAM, older features to disk; (4) limit satellite tile cache to 500MB in RAM, overflow to disk; (5) batch size 1 for all GPU operations; (6) explicit torch.cuda.empty_cache() between frames if VRAM pressure detected.
-
-### Confidence
-✅ High — standard memory management patterns.
-
---
-
-## Dimension 8: Security
-
-### Fact Confirmation
-JWT tokens are recommended for SSE endpoint security. API keys in query parameters are insecure (persist in logs, browser history).
-
-### Reference Comparison
-Draft01 mentions API key auth but no implementation details. SSE connections need proper authentication and resource limits.
-
-### Conclusion
-**Security improvements**: (1) JWT-based authentication for all endpoints; (2) short-lived tokens for SSE connections; (3) image folder whitelist (not just path traversal prevention — explicit whitelist of allowed base directories); (4) max concurrent SSE connections per client; (5) request rate limiting; (6) max image size validation; (7) all API keys in environment variables, never logged.
-
-### Confidence
-✅ High — standard security practices.
-
---
-
-## Dimension 9: Segment Management
-
-### Fact Confirmation
-According to Fact #8, YFS90 has re-localization capability after positioning failures. According to Fact #2, LightGlue may return false matches on non-overlapping pairs.
-
-### Reference Comparison
-Draft01's segment management relies on satellite matching to anchor each segment independently. If satellite matching fails, the segment stays "floating." No mechanism for cross-segment matching or delayed resolution.
-
-### Conclusion
-**Enhanced segment management**: (1) Explicit VO failure detection using match count + inlier ratio + geometric consistency (not just match count); (2) when a new segment gets satellite-anchored, attempt to connect to nearby floating segments using satellite-based position proximity; (3) DEM-based constraint: position must be consistent with terrain elevation; (4) configurable timeout for user input request — if no response within N frames, continue with best estimate and flag.
-
-### Confidence
-⚠️ Medium — cross-segment connection is logical but needs careful implementation to avoid false connections.
@@ -1,93 +0,0 @@
-# Validation Log — Solution Assessment (Mode B)
-
-## Validation Scenario 1: Normal flight over steppe with gradual turns
-
-**Scenario**: 1000-image flight over flat agricultural steppe. FullHD resolution. Starting GPS known. Gradual turns every 200 frames. Satellite imagery 2 years old.
-
-**Expected with Draft02 improvements**:
-1. XFeat VO processes frames at ~70ms each → well under 5s budget
-2. DINOv2 coarse retrieval finds correct satellite area despite 50-100m VO drift
-3. SuperPoint+LightGlue ONNX refines position to ~10-20m accuracy
-4. GTSAM iSAM2 smooths trajectory, reduces drift between anchors
-5. At gradual turns, VO continues working (overlap >30%)
-6. Processing stays under 1GB VRAM with 1600px downscale
-
-**Actual validation result**: Consistent with expectations. This is the "happy path" — both draft01 and draft02 would work. Draft02 advantage: faster processing, better optimizer.
-
-## Validation Scenario 2: Sharp turn with no overlap
-
-**Scenario**: After 500 normal frames, UAV makes a 90° sharp turn. Next 3 images have zero overlap with previous route. Then normal flight continues.
-
-**Expected with Draft02 improvements**:
-1. VO detects failure: match count drops below threshold → segment break
-2. LightGlue false-match protection: geometric consistency check rejects bad matches
-3. New segment starts. DINOv2 coarse retrieval searches wider area for satellite match
-4. If satellite match succeeds: new segment anchored, connected to previous via shared coordinate frame
-5. If satellite match fails: segment marked floating, user input requested (with timeout)
-6. After turn, if UAV returns near previous route, cross-segment connection attempted
-
-**Draft01 comparison**: Draft01 would also detect VO failure and create new segment, but lacks coarse retrieval → satellite matching depends entirely on VO estimate which may be wrong after turn. Higher risk of satellite match failure.
-
-## Validation Scenario 3: High-resolution images (6252×4168)
-
-**Scenario**: 500 images at full 6252×4168 resolution. RTX 2060 (6GB VRAM).
-
-**Expected with Draft02 improvements**:
-1. Images downscaled to 1600×1066 for feature extraction
-2. Full resolution preserved for GSD calculation only
-3. Per-frame VRAM: ~1.5GB for XFeat/SuperPoint + LightGlue
-4. RAM per frame: ~78MB raw + ~5MB features → manageable with streaming
-5. Total peak RAM: sliding window (50 frames × 5MB features) + satellite cache (500MB) + overhead ≈ 1.5GB pipeline
-6. Well within 16GB RAM budget
-
-**Actual validation result**: Consistent. Downscaling strategy is essential and was missing from draft01.
-
-## Validation Scenario 4: Outdated satellite imagery
-
-**Scenario**: Flight over area where Google Maps imagery is 4 years old. Significant changes: new buildings, removed forests, changed roads.
-
-**Expected with Draft02 improvements**:
-1. DINOv2 coarse retrieval: partial success (terrain structure still recognizable)
-2. SuperPoint+LightGlue fine matching: lower match count on changed areas
-3. Confidence score drops for affected frames → flagged in output
-4. Multi-provider fallback: try Mapbox tiles if Google matches are poor
-5. System falls back to VO-only for sections with no good satellite match
-6. User can provide custom satellite imagery for specific areas
-
-**Draft01 comparison**: Draft01 would also fail on changed areas but has no alternative provider and no coarse retrieval to help.
-
-## Validation Scenario 5: 3000-image flight hitting API rate limits
-
-**Scenario**: First flight in a new area. No cached tiles. 3000 images need ~2000 satellite tiles.
-
-**Expected with Draft02 improvements**:
-1. Initial download: 300 tiles around starting GPS (within rate limits)
-2. Progressive download as route extends: 5-20 tiles per frame
-3. Daily limit (15,000): sufficient for tiles but tight if multiple flights
-4. Request budgeting: prioritize tiles around current position, defer expansion
-5. Per-minute limit (6,000): no issue
-6. Monthly limit (100,000): covers ~50 flights at 2000 tiles each
-7. Mapbox fallback if Google budget exhausted
-
-**Draft01 comparison**: Draft01 assumed $200 free credit (expired). Rate limit analysis was incorrect.
-
-## Review Checklist
- [x] Draft conclusions consistent with fact cards
- [x] No important dimensions missed
- [x] No over-extrapolation
- [x] Conclusions actionable/verifiable
- [x] All scenarios plausible for the operational context
-
-## Counterexamples
- **Night flight**: Not addressed (out of scope — restriction says "mostly sunny weather")
- **Very low altitude (<100m)**: Satellite matching would have poor GSD match — not addressed but within restrictions (altitude ≤1km)
- **Urban area with tall buildings**: Homography VO degradation — mitigated by essential matrix fallback but not fully addressed
-
-## Conclusions Requiring No Revision
-All conclusions validated against scenarios. Key improvements are well-supported:
-1. Hierarchical satellite matching (coarse + fine)
-2. GTSAM factor graph optimization
-3. Multi-provider satellite tiles
-4. XFeat for VO speed
-5. Image downscaling for memory
-6. Proper security (JWT, rate limiting)
@@ -1,56 +0,0 @@
-# Question Decomposition
-
-## Original Question
-Assess current solution draft. Additionally:
-1. Try SuperPoint + LightGlue for visual odometry
-2. Can LiteSAM be SO SLOW because of big images? If we reduce size to 1280p, would that work faster?
-
-## Active Mode
-Mode B: Solution Assessment — `solution_draft01.md` exists in OUTPUT_DIR.
-
-## Question Type
-Problem Diagnosis + Decision Support
-
-## Research Subject Boundary
- **Population**: GPS-denied UAV navigation systems on edge hardware
- **Geography**: Eastern Ukraine conflict zone
- **Timeframe**: Current (2025-2026), using latest available tools
- **Level**: Jetson Orin Nano Super (8GB, 67 TOPS) — edge deployment
-
-## Decomposed Sub-Questions
-
-### Q1: SuperPoint + LightGlue for Visual Odometry
- What is SP+LG inference speed on Jetson-class hardware?
- How does it compare to cuVSLAM (116fps on Orin Nano)?
- Is SP+LG suitable for frame-to-frame VO at 3fps?
- What is SP+LG accuracy vs cuVSLAM for VO?
-
-### Q2: LiteSAM Speed vs Image Resolution
- What resolution was LiteSAM benchmarked at? (1184px on AGX Orin)
- How does LiteSAM speed scale with resolution?
- What would 1280px achieve on Orin Nano Super vs AGX Orin?
- Is the bottleneck image size or compute power gap?
-
-### Q3: General Weak Points in solution_draft01
- Are there functional weak points?
- Are there performance bottlenecks?
- Are there security gaps?
-
-### Q4: SP+LG for Satellite Matching (alternative to LiteSAM/XFeat)
- How does SP+LG perform on cross-view satellite-aerial matching?
- What does the LiteSAM paper say about SP+LG accuracy?
-
-## Timeliness Sensitivity Assessment
- **Research Topic**: Edge-deployed visual odometry and satellite-aerial matching
- **Sensitivity Level**: 🟠 High
- **Rationale**: cuVSLAM v15.0.0 released March 2026; LiteSAM published October 2025; LightGlue TensorRT optimizations actively evolving
- **Source Time Window**: 12 months
- **Priority official sources**:
-  1. LiteSAM paper (MDPI Remote Sensing, October 2025)
-  2. cuVSLAM / PyCuVSLAM v15.0.0 (March 2026)
-  3. LightGlue-ONNX / TensorRT benchmarks (2024-2026)
-  4. Intermodalics cuVSLAM benchmark (2025)
- **Key version information**:
-  - cuVSLAM: v15.0.0 (March 2026)
-  - LightGlue: ICCV 2023, TensorRT via fabio-sim/LightGlue-ONNX
-  - LiteSAM: Published October 2025, code at boyagesmile/LiteSAM
@@ -1,121 +0,0 @@
-# Source Registry
-
-## Source #1
- **Title**: LiteSAM: Lightweight and Robust Feature Matching for Satellite and Aerial Imagery
- **Link**: https://www.mdpi.com/2072-4292/17/19/3349
- **Tier**: L1
- **Publication Date**: 2025-10-01
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: LiteSAM v1.0; benchmarked on Jetson AGX Orin (JetPack 5.x era)
- **Target Audience**: UAV visual localization researchers and edge deployers
- **Research Boundary Match**: ✅ Full match
- **Summary**: LiteSAM (opt) achieves 497.49ms on Jetson AGX Orin at 1184px input. 6.31M params. RMSE@30 = 17.86m on UAV-VisLoc. Paper directly compares with SP+LG, stating "SP+LG achieves the fastest inference speed but at the expense of accuracy." Section 4.9 shows resolution vs speed tradeoff on RTX 3090Ti.
- **Related Sub-question**: Q2 (LiteSAM speed), Q4 (SP+LG for satellite matching)
-
-## Source #2
- **Title**: cuVSLAM: CUDA accelerated visual odometry and mapping
- **Link**: https://arxiv.org/abs/2506.04359
- **Tier**: L1
- **Publication Date**: 2025-06 (paper), v15.0.0 released 2026-03-10
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: cuVSLAM v15.0.0 / PyCuVSLAM v15.0.0
- **Target Audience**: Robotics/UAV visual odometry on NVIDIA Jetson
- **Research Boundary Match**: ✅ Full match
- **Summary**: CUDA-accelerated VO+SLAM, supports mono+IMU. 116fps on Jetson Orin Nano 8GB at 720p. <1% trajectory error on KITTI. <5cm on EuRoC.
- **Related Sub-question**: Q1 (SP+LG vs cuVSLAM)
-
-## Source #3
- **Title**: Intermodalics — NVIDIA Isaac ROS In-Depth: cuVSLAM and the DP3.1 Release
- **Link**: https://www.intermodalics.ai/blog/nvidia-isaac-ros-in-depth-cuvslam-and-the-dp3-1-release
- **Tier**: L2
- **Publication Date**: 2025 (DP3.1 release)
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: cuVSLAM v11 (DP3.1), benchmark data applicable to later versions
- **Target Audience**: Robotics developers using Isaac ROS
- **Research Boundary Match**: ✅ Full match
- **Summary**: 116fps on Orin Nano 8GB, 232fps on AGX Orin, 386fps on RTX 4060 Ti. Outperforms ORB-SLAM2 on KITTI.
- **Related Sub-question**: Q1
-
-## Source #4
- **Title**: Accelerating LightGlue Inference with ONNX Runtime and TensorRT
- **Link**: https://fabio-sim.github.io/blog/accelerating-lightglue-inference-onnx-runtime-tensorrt/
- **Tier**: L2
- **Publication Date**: 2024-07-17
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: torch 2.4.0, TensorRT 10.2.0, RTX 4080 benchmarks
- **Target Audience**: ML engineers deploying LightGlue
- **Research Boundary Match**: ⚠️ Partial (desktop GPU, not Jetson)
- **Summary**: TensorRT achieves 2-4x speedup over compiled PyTorch for SuperPoint+LightGlue. Full pipeline benchmarks on RTX 4080. TensorRT has 3840 keypoint limit. No Jetson-specific benchmarks provided.
- **Related Sub-question**: Q1
-
-## Source #5
- **Title**: LightGlue-with-FlashAttentionV2-TensorRT (Jetson Orin NX 8GB)
- **Link**: https://github.com/qdLMF/LightGlue-with-FlashAttentionV2-TensorRT
- **Tier**: L4
- **Publication Date**: 2025-02
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: TensorRT 8.5.2, Jetson Orin NX 8GB
- **Target Audience**: Edge ML deployers
- **Research Boundary Match**: ✅ Full match (similar hardware)
- **Summary**: CUTLASS-based FlashAttention V2 TensorRT plugin for LightGlue, tested on Jetson Orin NX 8GB. No published latency numbers, but confirms LightGlue TensorRT deployment on Orin-class hardware is feasible.
- **Related Sub-question**: Q1
-
-## Source #6
- **Title**: vo_lightglue — Visual Odometry with LightGlue
- **Link**: https://github.com/himadrir/vo_lightglue
- **Tier**: L4
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: N/A
- **Target Audience**: VO researchers
- **Research Boundary Match**: ⚠️ Partial (desktop, KITTI dataset)
- **Summary**: SP+LG achieves 10fps on KITTI dataset (desktop GPU). Odometric error ~1% vs 3.5-4.1% for FLANN-based matching. Much slower than cuVSLAM.
- **Related Sub-question**: Q1
-
-## Source #7
- **Title**: ForestVO: Enhancing Visual Odometry in Forest Environments through ForestGlue
- **Link**: https://arxiv.org/html/2504.01261v1
- **Tier**: L1
- **Publication Date**: 2025-04
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: N/A
- **Target Audience**: VO researchers
- **Research Boundary Match**: ⚠️ Partial (forest environment, not nadir UAV)
- **Summary**: SP+LG VO pipeline achieves 1.09m avg relative pose error, KITTI score 2.33%. Uses 512 keypoints (reduced from 2048) to cut compute. Outperforms DSO by 40%.
- **Related Sub-question**: Q1
-
-## Source #8
- **Title**: SuperPoint-SuperGlue-TensorRT (C++ deployment)
- **Link**: https://github.com/yuefanhao/SuperPoint-SuperGlue-TensorRT
- **Tier**: L4
- **Publication Date**: 2023-2024
- **Timeliness Status**: ⚠️ Needs verification (SuperGlue, not LightGlue)
- **Version Info**: TensorRT 8.x
- **Target Audience**: Edge deployers
- **Research Boundary Match**: ⚠️ Partial
- **Summary**: SuperPoint TensorRT extraction ~40ms on Jetson for 200 keypoints. C++ implementation.
- **Related Sub-question**: Q1
-
-## Source #9
- **Title**: Comparative Analysis of Advanced Feature Matching Algorithms in HSR Satellite Stereo
- **Link**: https://arxiv.org/abs/2405.06246
- **Tier**: L1
- **Publication Date**: 2024-05
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: N/A
- **Target Audience**: Remote sensing researchers
- **Research Boundary Match**: ⚠️ Partial (satellite stereo, not UAV-satellite cross-view)
- **Summary**: SP+LG shows "overall superior performance in balancing robustness, accuracy, distribution, and efficiency" for satellite stereo matching. But this is same-view satellite-satellite, not cross-view UAV-satellite.
- **Related Sub-question**: Q4
-
-## Source #10
- **Title**: PyCuVSLAM with reComputer (Seeed Studio)
- **Link**: https://wiki.seeedstudio.com/pycuvslam_recomputer_robotics/
- **Tier**: L3
- **Publication Date**: 2026
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: PyCuVSLAM v15.0.0, JetPack 6.2
- **Target Audience**: Robotics developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Tutorial for deploying PyCuVSLAM on Jetson Orin NX. Confirms mono+IMU mode, pip install from aarch64 wheel, EuRoC dataset examples.
- **Related Sub-question**: Q1
@@ -1,122 +0,0 @@
-# Fact Cards
-
-## Fact #1
- **Statement**: cuVSLAM achieves 116fps on Jetson Orin Nano 8GB at 720p resolution (~8.6ms/frame). 232fps on AGX Orin. 386fps on RTX 4060 Ti.
- **Source**: [Source #3] Intermodalics benchmark
- **Phase**: Assessment
- **Confidence**: ✅ High
- **Related Dimension**: VO speed comparison
-
-## Fact #2
- **Statement**: SuperPoint+LightGlue VO achieves ~10fps on KITTI dataset on desktop GPU (~100ms/frame). With 274 keypoints on RTX 2080Ti, LightGlue matching alone takes 33.9ms.
- **Source**: vo_lightglue, LG issue #36
- **Confidence**: ⚠️ Medium (desktop GPU, not Jetson)
- **Related Dimension**: VO speed comparison
-
-## Fact #3
- **Statement**: SuperPoint feature extraction takes ~40ms on Jetson (TensorRT, 200 keypoints).
- **Source**: SuperPoint-SuperGlue-TensorRT
- **Confidence**: ⚠️ Medium (older Jetson)
- **Related Dimension**: VO speed comparison
-
-## Fact #4
- **Statement**: LightGlue TensorRT with FlashAttention V2 has been deployed on Jetson Orin NX 8GB. No published latency numbers.
- **Source**: qdLMF/LightGlue-with-FlashAttentionV2-TensorRT
- **Confidence**: ⚠️ Medium
- **Related Dimension**: VO speed comparison
-
-## Fact #5
- **Statement**: LiteSAM (opt) inference: 61.98ms on RTX 3090, 497.49ms on Jetson AGX Orin at 1184px input. 6.31M params.
- **Source**: LiteSAM paper, abstract + Section 4.10
- **Confidence**: ✅ High
- **Related Dimension**: Satellite matcher speed
-
-## Fact #6
- **Statement**: Jetson AGX Orin has 275 TOPS INT8, 2048 CUDA cores. Orin Nano Super has 67 TOPS INT8, 1024 CUDA cores. AGX Orin is ~3-4x more powerful.
- **Source**: NVIDIA official specs
- **Confidence**: ✅ High
- **Related Dimension**: Hardware scaling
-
-## Fact #7
- **Statement**: LiteSAM processes at 1/8 scale internally. Coarse matching is O(N²) where N = (H/8 × W/8). For 1184px: ~21,904 tokens. For 1280px: ~25,600. For 480px: ~3,600.
- **Source**: LiteSAM paper, Sections 3.1-3.3
- **Confidence**: ✅ High
- **Related Dimension**: LiteSAM speed vs resolution
-
-## Fact #8
- **Statement**: LiteSAM paper Figure 1 states: "SP+LG achieves the fastest inference speed but at the expense of accuracy" vs LiteSAM on satellite-aerial benchmarks.
- **Source**: LiteSAM paper
- **Confidence**: ✅ High
- **Related Dimension**: SP+LG vs LiteSAM
-
-## Fact #9
- **Statement**: LiteSAM achieves RMSE@30 = 17.86m on UAV-VisLoc. SP+LG is worse on same benchmark.
- **Source**: LiteSAM paper
- **Confidence**: ✅ High
- **Related Dimension**: Satellite matcher accuracy
-
-## Fact #10
- **Statement**: cuVSLAM uses Shi-Tomasi corners ("Good Features to Track") for keypoint detection, divided into NxM grid patches. Uses Lucas-Kanade optical flow for tracking. When tracked keypoints fall below threshold, creates new keyframe.
- **Source**: cuVSLAM paper (arXiv:2506.04359), Section 2.1
- **Confidence**: ✅ High
- **Related Dimension**: cuVSLAM on difficult terrain
-
-## Fact #11
- **Statement**: cuVSLAM automatically switches to IMU when visual tracking fails (dark lighting, long solid surfaces). IMU integrator provides ~1 second of acceptable tracking. After IMU, constant-velocity integrator provides ~0.5 seconds more.
- **Source**: Isaac ROS cuVSLAM docs
- **Confidence**: ✅ High
- **Related Dimension**: cuVSLAM on difficult terrain
-
-## Fact #12
- **Statement**: cuVSLAM does NOT guarantee correct pose recovery after losing track. External algorithms required for global re-localization after tracking loss. Cannot fuse GNSS, wheel odometry, or LiDAR.
- **Source**: Intermodalics blog
- **Confidence**: ✅ High
- **Related Dimension**: cuVSLAM on difficult terrain
-
-## Fact #13
- **Statement**: cuVSLAM benchmarked on KITTI (mostly urban/suburban driving) and EuRoC (indoor drone). Neither benchmark includes nadir agricultural terrain, flat fields, or uniform vegetation. No published results for these conditions.
- **Source**: cuVSLAM paper Section 3
- **Confidence**: ✅ High
- **Related Dimension**: cuVSLAM on difficult terrain
-
-## Fact #14
- **Statement**: cuVSLAM multi-stereo mode "significantly improves accuracy and robustness on challenging sequences compared to single stereo cameras", designed for featureless surfaces (narrow corridors, elevators). But our system uses monocular camera only.
- **Source**: cuVSLAM paper Section 2.2.2
- **Confidence**: ✅ High
- **Related Dimension**: cuVSLAM on difficult terrain
-
-## Fact #15
- **Statement**: PFED achieves 97.15% Recall@1 on University-1652 at 251.5 FPS on AGX Orin with only 4.45G FLOPs. But this is image RETRIEVAL (which satellite tile matches), NOT pixel-level correspondence matching.
- **Source**: PFED paper (arXiv:2510.22582)
- **Confidence**: ✅ High
- **Related Dimension**: Satellite matching alternatives
-
-## Fact #16
- **Statement**: EfficientLoFTR is ~2.5x faster than LoFTR with higher accuracy. Semi-dense matcher, 15.05M params. Has TensorRT adaptation (LoFTR_TRT). Performs well on weak-texture areas where traditional methods fail. Designed for aerial imagery.
- **Source**: EfficientLoFTR paper (CVPR 2024), HuggingFace docs
- **Confidence**: ✅ High
- **Related Dimension**: Satellite matching alternatives
-
-## Fact #17
- **Statement**: Hierarchical AVL system (2025) uses two-stage approach: DINOv2 for coarse retrieval + SuperPoint for fine matching. 64.5-95% success rate on real-world drone trajectories. Includes IMU-based prior correction and sliding-window map updates.
- **Source**: MDPI Remote Sensing 2025
- **Confidence**: ✅ High
- **Related Dimension**: Satellite matching alternatives
-
-## Fact #18
- **Statement**: STHN uses deep homography estimation for UAV geo-localization: directly estimates homography transform (no feature detection/matching/RANSAC). Achieves 4.24m MACE at 50m range. Designed for thermal but architecture is modality-agnostic.
- **Source**: STHN paper (IEEE RA-L 2024)
- **Confidence**: ✅ High
- **Related Dimension**: Satellite matching alternatives
-
-## Fact #19
- **Statement**: For our nadir UAV → satellite matching, the cross-view gap is SMALL compared to typical cross-view problems (ground-to-satellite). Both views are approximately top-down. Main challenges: season/lighting, resolution mismatch, temporal changes. This means general-purpose matchers may work better than expected.
- **Source**: Analytical observation
- **Confidence**: ⚠️ Medium
- **Related Dimension**: Satellite matching alternatives
-
-## Fact #20
- **Statement**: LiteSAM paper benchmarked EfficientLoFTR (opt) on satellite-aerial: 19.8% slower than LiteSAM (opt) on AGX Orin but with 2.4x more parameters. EfficientLoFTR achieves competitive accuracy. LiteSAM paper Table 3/4 provides direct comparison.
- **Source**: LiteSAM paper, Section 4.5
- **Confidence**: ✅ High
- **Related Dimension**: EfficientLoFTR vs LiteSAM
@@ -1,45 +0,0 @@
-# Comparison Framework
-
-## Selected Framework Type
-Decision Support + Problem Diagnosis
-
-## Selected Dimensions
-1. Inference speed on Orin Nano Super
-2. Accuracy for the target task
-3. Cross-view robustness (satellite-aerial gap)
-4. Implementation complexity / ecosystem maturity
-5. Memory footprint
-6. TensorRT optimization readiness
-
-## Comparison 1: Visual Odometry — cuVSLAM vs SuperPoint+LightGlue
-
-| Dimension | cuVSLAM v15.0.0 | SuperPoint + LightGlue (TRT) | Factual Basis |
-|-----------|-----------------|-------------------------------|---------------|
-| Speed on Orin Nano | ~8.6ms/frame (116fps @ 720p) | Est. ~150-300ms/frame (SP ~40-60ms + LG ~100-200ms) | Fact #1, #2, #3 |
-| VO accuracy (KITTI) | <1% trajectory error | ~1% odometric error (desktop) | Fact #1, #2 |
-| VO accuracy (EuRoC) | <5cm position error | Not benchmarked | Fact #1 |
-| IMU integration | Native mono+IMU mode, auto-fallback | None — must add custom IMU fusion | Fact #1 |
-| Loop closure | Built-in | Not available | Fact #1 |
-| TensorRT ready | Native CUDA (not TensorRT, raw CUDA) | Requires ONNX export + TRT build | Fact #4 |
-| Memory | ~200-300MB | SP ~50MB + LG ~50-100MB = ~100-150MB | Fact #1 |
-| Implementation | pip install aarch64 wheel | Custom pipeline: SP export + LG export + matching + pose estimation | Fact #1, #4 |
-| Maturity on Jetson | NVIDIA-maintained, production-ready | Community TRT plugins, limited Jetson benchmarks | Fact #4, #5 |
-
-## Comparison 2: LiteSAM Speed at Different Resolutions
-
-| Dimension | 1184px (paper default) | 1280px (user proposal) | 640px | 480px | Factual Basis |
-|-----------|------------------------|------------------------|-------|-------|---------------|
-| Tokens at 1/8 scale | ~21,904 | ~25,600 | ~6,400 | ~3,600 | Fact #7 |
-| AGX Orin time | 497ms | Est. ~580ms (1.17x tokens) | Est. ~150ms | Est. ~90ms | Fact #5, #7 |
-| Orin Nano Super time (est.) | ~1.5-2.0s | ~1.7-2.3s | ~450-600ms | ~270-360ms | Fact #5, #6 |
-| Accuracy (RMSE@30) | 17.86m | Similar (slightly less) | Degraded | Significantly degraded | Fact #8, #10 |
-
-## Comparison 3: Satellite Matching — LiteSAM vs SP+LG vs XFeat
-
-| Dimension | LiteSAM (opt) | SuperPoint+LightGlue | XFeat semi-dense | Factual Basis |
-|-----------|--------------|---------------------|------------------|---------------|
-| Cross-view accuracy | RMSE@30 = 17.86m (UAV-VisLoc) | Worse than LiteSAM (paper confirms) | Not benchmarked on UAV-VisLoc | Fact #9, #10 |
-| Speed on Orin Nano (est.) | ~1.5-2s @ 1184px, ~270-360ms @ 480px | Est. ~100-200ms total | ~50-100ms | Fact #5, #2, existing draft |
-| Cross-view robustness | Designed for satellite-aerial gap | Sparse matcher, "lacks sufficient accuracy" for cross-view | General-purpose, less robust | Fact #9, #13 |
-| Parameters | 6.31M | SP ~1.3M + LG ~7M = ~8.3M | ~5M | Fact #5 |
-| Approach | Semi-dense (coarse-to-fine, subpixel) | Sparse (detect → match → verify) | Semi-dense (detect → KNN → refine) | Fact #1, existing draft |
@@ -1,90 +0,0 @@
-# Reasoning Chain
-
-## Dimension 1: SuperPoint+LightGlue for Visual Odometry
-
-### Fact Confirmation
-cuVSLAM achieves 116fps (~8.6ms/frame) on Orin Nano 8GB at 720p (Fact #1). SP+LG achieves ~10fps on KITTI on desktop GPU (Fact #2). SuperPoint alone takes ~40ms on Jetson for 200 keypoints (Fact #3). LightGlue matching on desktop GPU takes ~20-34ms for 274 keypoints (Fact #2).
-
-### Extrapolation to Orin Nano Super
-On Orin Nano Super, estimating SP+LG pipeline:
- SuperPoint extraction (1024 keypoints, 720p): ~50-80ms (based on Fact #3, scaled for more keypoints)
- LightGlue matching (TensorRT FP16, 1024 keypoints): ~80-200ms (based on Fact #11 — 2-4x speedup over PyTorch, but Orin Nano is ~4-6x slower than RTX 4080)
- Total SP+LG: ~130-280ms per frame
-
-cuVSLAM: ~8.6ms per frame.
-
-SP+LG would be **15-33x slower** than cuVSLAM for visual odometry on Orin Nano Super.
-
-### Additional Considerations
-cuVSLAM includes native IMU integration, loop closure, and auto-fallback. SP+LG provides none of these — they would need custom implementation, adding both development time and latency.
-
-### Conclusion
-**SP+LG is not viable as a cuVSLAM replacement for VO on Orin Nano Super.** cuVSLAM is purpose-built for Jetson and 15-33x faster. SP+LG's value lies in its accuracy for feature matching tasks, not real-time VO on edge hardware.
-
-### Confidence
-✅ High — performance gap is enormous and well-supported by multiple sources.
-
---
-
-## Dimension 2: LiteSAM Speed vs Image Resolution (1280px question)
-
-### Fact Confirmation
-LiteSAM (opt) achieves 497ms on AGX Orin at 1184px (Fact #5). AGX Orin is ~3-4x more powerful than Orin Nano Super (Fact #6). LiteSAM processes at 1/8 scale internally — coarse matching is O(N²) where N is proportional to resolution² (Fact #7).
-
-### Resolution Scaling Analysis
-
-**1280px vs 1184px**: Token count increases from ~21,904 to ~25,600 (+17%). Compute increases ~17-37% (linear to quadratic depending on bottleneck). This makes the problem WORSE, not better.
-
-**The user's intuition is likely**: "If 6252×4168 camera images are huge, maybe LiteSAM is slow because we feed it those big images. What if we use 1280px?" But the solution draft already specifies resizing to 480-640px before feeding LiteSAM. The 497ms benchmark on AGX Orin was already at 1184px (the UAV-VisLoc benchmark resolution).
-
-**The real bottleneck is hardware, not image size:**
- At 1184px on AGX Orin: 497ms → on Orin Nano Super: est. **~1.5-2.0s**
- At 1280px on Orin Nano Super: est. **~1.7-2.3s** (WORSE — more tokens)
- At 640px on Orin Nano Super: est. **~450-600ms** (borderline)
- At 480px on Orin Nano Super: est. **~270-360ms** (possibly within 400ms budget)
-
-### Conclusion
-**1280px would make LiteSAM SLOWER, not faster.** The paper benchmarked at 1184px. The bottleneck is the hardware gap (AGX Orin 275 TOPS → Orin Nano Super 67 TOPS). To make LiteSAM fit the 400ms budget, resolution must drop to ~480px, which may significantly degrade cross-view matching accuracy. The original solution draft's approach (benchmark at 480px, abandon if too slow) remains correct.
-
-### Confidence
-✅ High — paper benchmarks + hardware specs provide strong basis.
-
---
-
-## Dimension 3: SP+LG for Satellite Matching (alternative to LiteSAM)
-
-### Fact Confirmation
-LiteSAM paper explicitly states "SP+LG achieves the fastest inference speed but at the expense of accuracy" on satellite-aerial benchmarks (Fact #9). SP+LG is a sparse matcher; the paper notes sparse matchers "lack sufficient accuracy" for cross-view UAV-satellite matching due to texture-scarce regions (Fact #13). LiteSAM achieves RMSE@30 = 17.86m; SP+LG is worse (Fact #10).
-
-### Speed Advantage of SP+LG
-On Orin Nano Super, SP+LG satellite matching pipeline:
- SuperPoint extraction (both images): ~50-80ms × 2 images
- LightGlue matching: ~80-200ms
- Total: ~180-360ms
-
-This is competitive with the 400ms budget. But accuracy is worse than LiteSAM.
-
-### Comparison with XFeat
-XFeat semi-dense: ~50-100ms on Orin Nano Super (from existing draft). XFeat is 2-4x faster than SP+LG and also handles semi-dense matching. For the satellite matching role, XFeat is a better "fast fallback" than SP+LG.
-
-### Conclusion
-**SP+LG is not recommended for satellite matching.** It's slower than XFeat and less accurate than LiteSAM for cross-view matching. XFeat remains the better fallback. SP+LG could serve as a third-tier fallback, but the added complexity isn't justified given XFeat's advantages.
-
-### Confidence
-✅ High — direct comparison from the LiteSAM paper.
-
---
-
-## Dimension 4: Other Weak Points in solution_draft01
-
-### cuVSLAM Nadir Camera Concern
-The solution correctly flags cuVSLAM's "nadir-only camera" as untested. cuVSLAM was designed for robotics (forward-facing cameras). Nadir UAV camera looking straight down at terrain has different motion characteristics. However, cuVSLAM supports arbitrary camera configurations and IMU mode should compensate. **Risk is MEDIUM, mitigation is adequate** (XFeat fallback).
-
-### Memory Budget Gap
-The solution estimates ~1.9-2.4GB total. This looks optimistic if cuVSLAM needs to maintain a map for loop closure. The cuVSLAM map grows over time. For a 3000-frame flight (~16 min at 3fps), map memory could grow to 500MB-1GB. **Risk: memory pressure late in flight.** Mitigation: configure cuVSLAM map pruning, limit map size.
-
-### Tile Search Strategy Underspecified
-The solution mentions GeoHash-indexed tiles but doesn't detail how the system determines which tile to match against when ESKF position has high uncertainty (e.g., after VO failure). The expanded search (±1km) could require loading 10-20 tiles, which is slow from storage.
-
-### Confidence
-⚠️ Medium — these are analytical observations, not empirically verified.
@@ -1,52 +0,0 @@
-# Validation Log
-
-## Validation Scenario 1: SP+LG for VO during Normal Flight
-
-A UAV flies straight at 3fps. Each frame needs VO within 400ms.
-
-### Expected Based on Conclusions
-cuVSLAM: processes each frame in ~8.6ms, leaves 391ms for satellite matching and fusion. Immediate VO result via SSE.
-SP+LG: processes each frame in ~130-280ms, leaves ~120-270ms. May interfere with satellite matching CUDA resources.
-
-### Actual Validation
-cuVSLAM is clearly superior. SP+LG offers no advantage here — cuVSLAM is 15-33x faster AND includes IMU fallback. SP+LG would require building a custom VO pipeline around a feature matcher, whereas cuVSLAM is a complete VO solution.
-
-### Counterexamples
-If cuVSLAM fails on nadir camera (its main risk), SP+LG could serve as a fallback VO method. But XFeat frame-to-frame (~30-50ms) is already identified as the cuVSLAM fallback and is 3-6x faster than SP+LG.
-
-## Validation Scenario 2: LiteSAM at 1280px on Orin Nano Super
-
-A keyframe needs satellite matching. Image is resized to 1280px for LiteSAM.
-
-### Expected Based on Conclusions
-LiteSAM at 1280px on Orin Nano Super: ~1.7-2.3s. This is 4-6x over the 400ms budget. Even running async, it means satellite corrections arrive 5-7 frames later.
-
-### Actual Validation
-1280px is LARGER than the paper's 1184px benchmark resolution. The user likely assumed we feed the full camera image (6252×4168) to LiteSAM, causing slowness. But the solution already downsamples. The bottleneck is the hardware performance gap (Orin Nano Super = ~25% of AGX Orin compute).
-
-### Counterexamples
-If LiteSAM's TensorRT FP16 engine with reparameterized MobileOne achieves better optimization than the paper's AMP benchmark (which uses PyTorch, not TensorRT), speed could improve 2-3x. At 480px with TensorRT FP16: potentially ~90-180ms on Orin Nano Super. This is worth benchmarking.
-
-## Validation Scenario 3: SP+LG as Satellite Matcher After LiteSAM Abandonment
-
-LiteSAM fails benchmark. Instead of XFeat, we try SP+LG for satellite matching.
-
-### Expected Based on Conclusions
-SP+LG: ~180-360ms on Orin Nano Super. Accuracy is worse than LiteSAM for cross-view matching.
-XFeat: ~50-100ms. Accuracy is unproven on cross-view but general-purpose semi-dense.
-
-### Actual Validation
-SP+LG is 2-4x slower than XFeat and the LiteSAM paper confirms worse accuracy for satellite-aerial. XFeat's semi-dense approach is more suited to the texture-scarce regions common in UAV imagery. SP+LG's sparse keypoint detection may fail on agricultural fields or water bodies.
-
-### Counterexamples
-SP+LG could outperform XFeat on high-texture urban areas where sparse features are abundant. But the operational region (eastern Ukraine) is primarily agricultural, making this advantage unlikely.
-
-## Review Checklist
- [x] Draft conclusions consistent with fact cards
- [x] No important dimensions missed
- [x] No over-extrapolation
- [x] Conclusions actionable/verifiable
- [ ] Note: Orin Nano Super estimates are extrapolated from AGX Orin data using the 3-4x compute ratio. Day-one benchmarking remains essential.
-
-## Conclusions Requiring Revision
-None — the original solution draft's architecture (cuVSLAM for VO, benchmark-driven LiteSAM/XFeat for satellite) is confirmed sound. SP+LG is not recommended for either role on this hardware.
@@ -1,102 +0,0 @@
-# Question Decomposition
-
-## Original Question
-Assess solution_draft02.md against updated acceptance criteria and restrictions. The AC and restrictions have been significantly revised to reflect real onboard deployment requirements (MAVLink integration, ground station telemetry, startup/failsafe, object localization, thermal management, satellite imagery specs).
-
-## Active Mode
-Mode B: Solution Assessment — `solution_draft02.md` is the latest draft in OUTPUT_DIR.
-
-## Question Type
-Problem Diagnosis + Decision Support
-
-## Research Subject Boundary
- **Population**: GPS-denied UAV navigation systems on edge hardware (Jetson Orin Nano Super)
- **Geography**: Eastern/southern Ukraine (east of Dnipro River), conflict zone
- **Timeframe**: Current (2025-2026), latest available tools and libraries
- **Level**: Onboard companion computer, real-time flight controller integration via MAVLink
-
-## Key Delta: What Changed in AC/Restrictions
-
-### Restrictions Changes
-1. Two cameras: Navigation (fixed, downward) + AI camera (configurable angle/zoom)
-2. Processing on Jetson Orin Nano Super (was "stationary computer or laptop")
-3. IMU data IS available via flight controller (was "NO data from IMU")
-4. MAVLink protocol via MAVSDK for flight controller communication
-5. Must output GPS_INPUT messages as GPS replacement
-6. Ground station telemetry link available but bandwidth-limited
-7. Thermal throttling must be accounted for
-8. Satellite imagery pre-loaded, storage limited
-
-### Acceptance Criteria Changes
-1. <400ms per frame to flight controller (was <5s for processing)
-2. MAVLink GPS_INPUT output (was REST API + SSE)
-3. Ground station: stream position/confidence, receive re-localization commands
-4. Object localization: trigonometric GPS from AI camera angle/zoom/altitude
-5. Startup: initialize from last known GPS before GPS denial
-6. Failsafe: IMU-only fallback after N seconds of total failure
-7. Reboot recovery: re-initialize from flight controller IMU-extrapolated position
-8. Max cumulative VO drift <100m between satellite anchors
-9. Confidence score per position estimate (high/low)
-10. Satellite imagery: ≥0.5 m/pixel, <2 years old
-11. WGS84 output format
-12. Re-localization via telemetry to ground station (not REST API user input)
-
-## Decomposed Sub-Questions
-
-### Q1: MAVLink GPS_INPUT Integration
- How does MAVSDK Python handle GPS_INPUT messages?
- What fields are required in GPS_INPUT?
- What update rate does the flight controller expect?
- Can we send confidence/accuracy indicators via MAVLink?
- How does this replace the REST API + SSE architecture?
-
-### Q2: Ground Station Telemetry Integration
- How to stream position + confidence over bandwidth-limited telemetry?
- How to receive operator re-localization commands?
- What MAVLink messages support custom data?
- What bandwidth is typical for UAV telemetry links?
-
-### Q3: Startup & Failsafe Mechanisms
- How to initialize from flight controller's last GPS position?
- How to detect GPS denial onset?
- What happens on companion computer reboot mid-flight?
- How to implement IMU-only dead reckoning fallback?
-
-### Q4: Object Localization via AI Camera
- How to compute ground GPS from UAV position + camera angle + zoom + altitude?
- What accuracy can be expected given GPS-denied position error?
- How to handle the API between GPS-denied system and AI detection system?
-
-### Q5: Thermal Management on Jetson Orin Nano Super
- What is sustained thermal performance under GPU load?
- How to monitor and mitigate thermal throttling?
- What power modes are available?
-
-### Q6: VO Drift Budget & Monitoring
- How to measure cumulative drift between satellite anchors?
- How to trigger satellite matching when drift approaches 100m?
- ESKF covariance as drift proxy?
-
-### Q7: Weak Points in Draft02 Architecture
- REST API + SSE architecture is wrong — must be MAVLink
- No ground station integration
- No startup/shutdown procedures
- No thermal management
- No object localization detail for AI camera with configurable angle/zoom
- Memory budget doesn't account for MAVSDK overhead
-
-## Timeliness Sensitivity Assessment
- **Research Topic**: MAVLink integration, MAVSDK for Jetson, ground station telemetry, thermal management
- **Sensitivity Level**: 🟠 High
- **Rationale**: MAVSDK actively developed; MAVLink message set evolving; Jetson JetPack 6.2 specific
- **Source Time Window**: 12 months
- **Priority official sources**:
-  1. MAVSDK Python documentation (mavsdk.io)
-  2. MAVLink message definitions (mavlink.io)
-  3. NVIDIA Jetson Orin Nano thermal documentation
-  4. PX4/ArduPilot GPS_INPUT documentation
- **Key version information**:
-  - MAVSDK-Python: latest PyPI version
-  - MAVLink: v2 protocol
-  - JetPack: 6.2.2
-  - PyCuVSLAM: v15.0.0
@@ -1,175 +0,0 @@
-# Source Registry
-
-## Source #1
- **Title**: MAVSDK-Python Issue #320: Input external GPS through MAVSDK
- **Link**: https://github.com/mavlink/MAVSDK-Python/issues/320
- **Tier**: L4
- **Publication Date**: 2021 (still open 2025)
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: MAVSDK-Python — GPS_INPUT not supported as of v3.15.3
- **Target Audience**: Companion computer developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: MAVSDK-Python does not support GPS_INPUT message. Feature requested but unresolved.
- **Related Sub-question**: Q1
-
-## Source #2
- **Title**: MAVLink GPS_INPUT Message Specification (mavlink_msg_gps_input.h)
- **Link**: https://rflysim.com/doc/en/RflySimAPIs/RflySimSDK/html/mavlink__msg__gps__input_8h_source.html
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: MAVLink v2, Message ID 232
- **Target Audience**: MAVLink developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: GPS_INPUT message fields: lat/lon (1E7), alt, fix_type, horiz_accuracy, vert_accuracy, speed_accuracy, hdop, vdop, satellites_visible, velocities NED, yaw, ignore_flags.
- **Related Sub-question**: Q1
-
-## Source #3
- **Title**: ArduPilot GPS Input MAVProxy Documentation
- **Link**: https://ardupilot.org/mavproxy/docs/modules/GPSInput.html
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: ArduPilot GPS1_TYPE=14
- **Target Audience**: ArduPilot companion computer developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: GPS_INPUT requires GPS1_TYPE=14. Accepts JSON over UDP port 25100. Fields: lat, lon, alt, fix_type, hdop, timestamps.
- **Related Sub-question**: Q1
-
-## Source #4
- **Title**: pymavlink GPS_INPUT example
- **Link**: https://webperso.ensta.fr/lebars/Share/GPS_INPUT_pymavlink.py
- **Tier**: L3
- **Publication Date**: 2023
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: pymavlink
- **Target Audience**: Companion computer developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Complete pymavlink example for sending GPS_INPUT with all fields including yaw. Uses gps_input_send() method.
- **Related Sub-question**: Q1
-
-## Source #5
- **Title**: ArduPilot AP_GPS_Params.cpp — GPS_RATE_MS
- **Link**: https://cocalc.com/github/ardupilot/ardupilot/blob/master/libraries/AP_GPS/AP_GPS_Params.cpp
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: ArduPilot master
- **Target Audience**: ArduPilot developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: GPS_RATE_MS default 200ms (5Hz), range 50-200ms (5-20Hz). Below 5Hz not allowed.
- **Related Sub-question**: Q1
-
-## Source #6
- **Title**: MAVLink Telemetry Bandwidth Optimization Issue #1605
- **Link**: https://github.com/mavlink/mavlink/issues/1605
- **Tier**: L2
- **Publication Date**: 2022
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: MAVLink protocol developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Minimal telemetry requires ~12kbit/s. Optimized ~6kbit/s. SiK at 57600 baud provides ~21% usable budget. RFD900 for long range (15km+).
- **Related Sub-question**: Q2
-
-## Source #7
- **Title**: NVIDIA JetPack 6.2 Super Mode Blog
- **Link**: https://developer.nvidia.com/blog/nvidia-jetpack-6-2-brings-super-mode-to-nvidia-jetson-orin-nano-and-jetson-orin-nx-modules/
- **Tier**: L1
- **Publication Date**: 2025-01
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: JetPack 6.2, Orin Nano Super
- **Target Audience**: Jetson developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: MAXN SUPER mode for peak performance. Thermal throttling at 80°C. Power modes: 15W, 25W, MAXN SUPER. Up to 1.7x AI boost.
- **Related Sub-question**: Q5
-
-## Source #8
- **Title**: Jetson Orin Nano Power Consumption Analysis
- **Link**: https://edgeaistack.app/blog/jetson-orin-nano-power-consumption/
- **Tier**: L3
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Jetson edge deployment engineers
- **Research Boundary Match**: ✅ Full match
- **Summary**: 5W idle, 8-12W typical inference, 25W peak. Throttling above 80°C drops GPU from 1GHz to 300MHz. Active cooling required for sustained loads.
- **Related Sub-question**: Q5
-
-## Source #9
- **Title**: UAV Target Geolocation (Sensors 2022)
- **Link**: https://www.mdpi.com/1424-8220/22/5/1903
- **Tier**: L1
- **Publication Date**: 2022
- **Timeliness Status**: ✅ Currently valid (math doesn't change)
- **Target Audience**: UAV reconnaissance systems
- **Research Boundary Match**: ✅ Full match
- **Summary**: Trigonometric target geolocation from camera angle, altitude, UAV position. Iterative refinement improves accuracy 22-38x.
- **Related Sub-question**: Q4
-
-## Source #10
- **Title**: pymavlink vs MAVSDK-Python for custom messages (Issue #739)
- **Link**: https://github.com/mavlink/MAVSDK-Python/issues/739
- **Tier**: L4
- **Publication Date**: 2024-12
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: MAVSDK-Python, pymavlink
- **Target Audience**: Companion computer developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: MAVSDK-Python lacks custom message support. pymavlink recommended for GPS_INPUT and custom messages. MAVSDK v4 may add MavlinkDirect plugin.
- **Related Sub-question**: Q1
-
-## Source #11
- **Title**: NAMED_VALUE_FLOAT for custom telemetry (PR #18501)
- **Link**: https://github.com/ArduPilot/ardupilot/pull/18501
- **Tier**: L2
- **Publication Date**: 2022
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: ArduPilot companion computer developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: NAMED_VALUE_FLOAT messages from companion computer are logged by ArduPilot and forwarded to GCS. Useful for custom telemetry data.
- **Related Sub-question**: Q2
-
-## Source #12
- **Title**: ArduPilot Companion Computer UART Connection
- **Link**: https://ardupilot.org/dev/docs/raspberry-pi-via-mavlink.html
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: ArduPilot companion computer developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Connect via TELEM2 UART. SERIAL2_PROTOCOL=2 (MAVLink2). Baud up to 1.5Mbps. TX/RX crossover.
- **Related Sub-question**: Q1, Q2
-
-## Source #13
- **Title**: Jetson Orin Nano UART with ArduPilot
- **Link**: https://forums.developer.nvidia.com/t/uart-connection-between-jetson-nano-orin-and-ardupilot/325416
- **Tier**: L4
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: JetPack 6.x, Orin Nano
- **Target Audience**: Jetson + ArduPilot integration
- **Research Boundary Match**: ✅ Full match
- **Summary**: UART instability reported on Orin Nano with ArduPilot. Use /dev/ttyTHS0 or /dev/ttyTHS1. Check pinout carefully.
- **Related Sub-question**: Q1
-
-## Source #14
- **Title**: MAVSDK-Python v3.15.3 PyPI (aarch64 wheels)
- **Link**: https://pypi.org/project/mavsdk/
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: v3.15.3, manylinux2014_aarch64
- **Target Audience**: MAVSDK Python developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: MAVSDK-Python has aarch64 wheels. pip install works on Jetson. But no GPS_INPUT support.
- **Related Sub-question**: Q1
-
-## Source #15
- **Title**: ArduPilot receive COMMAND_LONG on companion computer
- **Link**: https://discuss.ardupilot.org/t/recieve-mav-cmd-on-companion-computer/48928
- **Tier**: L4
- **Publication Date**: 2020
- **Timeliness Status**: ⚠️ Needs verification (old but concept still valid)
- **Target Audience**: ArduPilot companion computer developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Companion computer can receive COMMAND_LONG messages from GCS via MAVLink. ArduPilot scripting can intercept specific command IDs.
- **Related Sub-question**: Q2
@@ -1,105 +0,0 @@
-# Fact Cards
-
-## Fact #1
- **Statement**: MAVSDK-Python (v3.15.3) does NOT support sending GPS_INPUT MAVLink messages. The feature has been requested since 2021 and remains unresolved. Custom message support is planned for MAVSDK v4 but not available in Python wrapper.
- **Source**: Source #1, #10, #14
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — confirmed by MAVSDK maintainers
- **Related Dimension**: Flight Controller Integration
-
-## Fact #2
- **Statement**: pymavlink provides full access to all MAVLink messages including GPS_INPUT via `mav.gps_input_send()`. It is the recommended library for companion computers that need to send GPS_INPUT messages.
- **Source**: Source #4, #10
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — working examples exist
- **Related Dimension**: Flight Controller Integration
-
-## Fact #3
- **Statement**: GPS_INPUT (MAVLink msg ID 232) contains: lat/lon (WGS84, degrees×1E7), alt (AMSL), fix_type (0-8), horiz_accuracy (m), vert_accuracy (m), speed_accuracy (m/s), hdop, vdop, satellites_visible, vn/ve/vd (NED m/s), yaw (centidegrees), gps_id, ignore_flags.
- **Source**: Source #2
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — official MAVLink spec
- **Related Dimension**: Flight Controller Integration
-
-## Fact #4
- **Statement**: ArduPilot requires GPS1_TYPE=14 (MAVLink) to accept GPS_INPUT messages from a companion computer. Connection via TELEM2 UART, SERIAL2_PROTOCOL=2 (MAVLink2).
- **Source**: Source #3, #12
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — official ArduPilot documentation
- **Related Dimension**: Flight Controller Integration
-
-## Fact #5
- **Statement**: ArduPilot GPS update rate (GPS_RATE_MS) default is 200ms (5Hz), range 50-200ms (5-20Hz). Our camera at 3fps (333ms) means GPS_INPUT at 3Hz. ArduPilot minimum is 5Hz. We must interpolate/predict between camera frames to meet 5Hz minimum.
- **Source**: Source #5
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — ArduPilot source code
- **Related Dimension**: Flight Controller Integration
-
-## Fact #6
- **Statement**: GPS_INPUT horiz_accuracy field directly maps to our confidence scoring. We can report: satellite-anchored ≈ 10-20m accuracy, VO-extrapolated ≈ 20-50m, IMU-only ≈ 100m+. ArduPilot EKF uses this for fusion weighting internally.
- **Source**: Source #2, #3
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ⚠️ Medium — accuracy mapping is estimated, EKF weighting not fully documented
- **Related Dimension**: Flight Controller Integration
-
-## Fact #7
- **Statement**: Typical UAV telemetry bandwidth: SiK radio at 57600 baud provides ~12kbit/s usable for MAVLink. RFD900 provides long range (15km+) at similar data rates. Position telemetry must be compact — ~50 bytes per position update.
- **Source**: Source #6
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — MAVLink protocol specs
- **Related Dimension**: Ground Station Telemetry
-
-## Fact #8
- **Statement**: NAMED_VALUE_FLOAT MAVLink message can stream custom telemetry from companion computer to ground station. ArduPilot logs and forwards these. Mission Planner displays them. Useful for confidence score, drift status, matching status.
- **Source**: Source #11
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — ArduPilot merged PR
- **Related Dimension**: Ground Station Telemetry
-
-## Fact #9
- **Statement**: Jetson Orin Nano Super throttles GPU from 1GHz to ~300MHz when junction temperature exceeds 80°C. Active cooling (fan) required for sustained load. Power consumption: 5W idle, 8-12W typical inference, 25W peak. Modes: 15W, 25W, MAXN SUPER.
- **Source**: Source #7, #8
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — NVIDIA official
- **Related Dimension**: Thermal Management
-
-## Fact #10
- **Statement**: Jetson Orin Nano UART connection to ArduPilot uses /dev/ttyTHS0 or /dev/ttyTHS1. UART instability reported on some units — verify pinout, use JetPack 6.2.2+. Baud up to 1.5Mbps supported.
- **Source**: Source #12, #13
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ⚠️ Medium — UART instability is a known issue with workarounds
- **Related Dimension**: Flight Controller Integration
-
-## Fact #11
- **Statement**: Object geolocation from UAV: for nadir (downward) camera, pixel offset from center → meters via GSD → rotate by heading → add to UAV GPS. For oblique (AI) camera with angle θ from vertical: ground_distance = altitude × tan(θ). Combined with zoom → effective focal length → pixel-to-meter conversion. Flat terrain assumption simplifies to basic trigonometry.
- **Source**: Source #9
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — well-established trigonometry
- **Related Dimension**: Object Localization
-
-## Fact #12
- **Statement**: Companion computer can receive COMMAND_LONG from ground station via MAVLink. For re-localization hints: ground station sends a custom command with approximate lat/lon, companion computer receives it via pymavlink message listener.
- **Source**: Source #15
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ⚠️ Medium — specific implementation for re-localization hint would be custom
- **Related Dimension**: Ground Station Telemetry
-
-## Fact #13
- **Statement**: The restrictions.md now says "using MAVSDK library" but MAVSDK-Python cannot send GPS_INPUT. pymavlink is the only viable Python option for GPS_INPUT. This is a restriction conflict that must be resolved — use pymavlink for GPS_INPUT (core function) or accept MAVSDK + pymavlink hybrid.
- **Source**: Source #1, #2, #10
- **Phase**: Assessment
- **Target Audience**: GPS-Denied system developers
- **Confidence**: ✅ High — confirmed limitation
- **Related Dimension**: Flight Controller Integration
@@ -1,62 +0,0 @@
-# Comparison Framework
-
-## Selected Framework Type
-Problem Diagnosis + Decision Support (Mode B)
-
-## Weak Point Dimensions (from draft02 → new AC/restrictions)
-
-### Dimension 1: Output Architecture (CRITICAL)
-Draft02 uses FastAPI + SSE to stream positions to clients.
-New AC requires MAVLink GPS_INPUT to flight controller as PRIMARY output.
-Entire output architecture must change.
-
-### Dimension 2: Ground Station Communication (CRITICAL)
-Draft02 has no ground station integration.
-New AC requires: stream position+confidence via telemetry, receive re-localization commands.
-
-### Dimension 3: MAVLink Library Choice (CRITICAL)
-Restrictions say "MAVSDK library" but MAVSDK-Python cannot send GPS_INPUT.
-Must use pymavlink for core function.
-
-### Dimension 4: GPS Update Rate (HIGH)
-Camera at 3fps → 3Hz position updates. ArduPilot minimum GPS rate is 5Hz.
-Need IMU-based interpolation between camera frames.
-
-### Dimension 5: Startup & Failsafe (HIGH)
-Draft02 has no initialization or failsafe procedures.
-New AC requires: init from last GPS, reboot recovery, IMU fallback after N seconds.
-
-### Dimension 6: Object Localization (MEDIUM)
-Draft02 has basic pixel-to-GPS for navigation camera only.
-New AC adds AI camera with configurable angle, zoom — trigonometric projection needed.
-
-### Dimension 7: Thermal Management (MEDIUM)
-Draft02 ignores thermal throttling.
-Jetson Orin Nano Super throttles at 80°C — can drop GPU 3x.
-
-### Dimension 8: VO Drift Budget Monitoring (MEDIUM)
-New AC: max cumulative VO drift <100m between satellite anchors.
-Draft02 uses ESKF covariance but doesn't explicitly track drift budget.
-
-### Dimension 9: Satellite Imagery Specs (LOW)
-New AC: ≥0.5 m/pixel, <2 years old. Draft02 uses Google Maps zoom 18-19 which is ~0.3-0.6 m/pixel.
-Mostly compatible, needs explicit validation.
-
-### Dimension 10: API for Internal Systems (LOW)
-Object localization requests from AI systems need a local IPC mechanism.
-FastAPI could be retained for local-only inter-process communication.
-
-## Initial Population
-
-| Dimension | Draft02 State | Required State | Gap Severity |
-|-----------|--------------|----------------|-------------|
-| Output Architecture | FastAPI + SSE to client | MAVLink GPS_INPUT to flight controller | CRITICAL — full redesign |
-| Ground Station | None | Bidirectional MAVLink telemetry | CRITICAL — new component |
-| MAVLink Library | Not applicable (no MAVLink) | pymavlink (MAVSDK can't do GPS_INPUT) | CRITICAL — new dependency |
-| GPS Update Rate | 3fps → ~3Hz output | ≥5Hz to ArduPilot EKF | HIGH — need IMU interpolation |
-| Startup & Failsafe | None | Init from GPS, reboot recovery, IMU fallback | HIGH — new procedures |
-| Object Localization | Basic nadir pixel-to-GPS | AI camera angle+zoom trigonometry | MEDIUM — extend existing |
-| Thermal Management | Not addressed | Monitor + mitigate throttling | MEDIUM — operational concern |
-| VO Drift Budget | ESKF covariance only | Explicit <100m tracking + trigger | MEDIUM — extend ESKF |
-| Satellite Imagery Specs | Google Maps zoom 18-19 | ≥0.5 m/pixel, <2 years | LOW — mostly met |
-| Internal IPC | REST API | Lightweight local API or shared memory | LOW — simplify from draft02 |
@@ -1,202 +0,0 @@
-# Reasoning Chain
-
-## Dimension 1: Output Architecture
-
-### Fact Confirmation
-Per Fact #3, GPS_INPUT (MAVLink msg ID 232) accepts lat/lon in WGS84 (degrees×1E7), altitude, fix_type, accuracy fields, and NED velocities. Per Fact #4, ArduPilot uses GPS1_TYPE=14 to accept MAVLink GPS input. The flight controller's EKF fuses this as if it were a real GPS module.
-
-### Reference Comparison
-Draft02 uses FastAPI + SSE to stream position data to a REST client. The new AC requires the system to output GPS coordinates directly to the flight controller via MAVLink GPS_INPUT, replacing the real GPS module. The flight controller then uses these coordinates for navigation/autopilot functions. The ground station receives position data indirectly via the flight controller's telemetry forwarding.
-
-### Conclusion
-The entire output architecture must change from REST API + SSE → pymavlink GPS_INPUT sender. FastAPI is no longer the primary output mechanism. It may be retained only for local IPC with other onboard AI systems (object localization requests). The SSE streaming to external clients is replaced by MAVLink telemetry forwarding through the flight controller.
-
-### Confidence
-✅ High — clear requirement change backed by MAVLink specification
-
---
-
-## Dimension 2: Ground Station Communication
-
-### Fact Confirmation
-Per Fact #7, typical telemetry bandwidth is ~12kbit/s (SiK). Per Fact #8, NAMED_VALUE_FLOAT can stream custom values from companion to GCS. Per Fact #12, COMMAND_LONG can deliver commands from GCS to companion.
-
-### Reference Comparison
-Draft02 has no ground station integration. The new AC requires:
-1. Stream position + confidence to ground station (passive, via telemetry forwarding of GPS_INPUT data + custom NAMED_VALUE_FLOAT for confidence/drift)
-2. Receive re-localization commands from operator (active, via COMMAND_LONG or custom MAVLink message)
-
-### Conclusion
-Ground station communication uses MAVLink messages forwarded through the flight controller's telemetry radio. Position data flows automatically (flight controller forwards GPS data to GCS). Custom telemetry (confidence, drift, status) uses NAMED_VALUE_FLOAT. Re-localization hints from operator use a custom COMMAND_LONG with lat/lon payload. Bandwidth is tight (~12kbit/s) so minimize custom message frequency (1-2Hz max for NAMED_VALUE_FLOAT).
-
-### Confidence
-✅ High — standard MAVLink patterns
-
---
-
-## Dimension 3: MAVLink Library Choice
-
-### Fact Confirmation
-Per Fact #1, MAVSDK-Python v3.15.3 does NOT support GPS_INPUT. Per Fact #2, pymavlink provides full GPS_INPUT support via `mav.gps_input_send()`. Per Fact #13, the restrictions say "using MAVSDK library" but MAVSDK literally cannot do the core function.
-
-### Reference Comparison
-MAVSDK is a higher-level abstraction over MAVLink. pymavlink is the lower-level direct MAVLink implementation. For GPS_INPUT (our core output), only pymavlink works.
-
-### Conclusion
-Use **pymavlink** as the MAVLink library. The restriction mentioning MAVSDK must be noted as a conflict — pymavlink is the only viable option for GPS_INPUT in Python. pymavlink is lightweight, pure Python, works on aarch64, and provides full access to all MAVLink messages. MAVSDK v4 may add custom message support in the future but is not available now.
-
-### Confidence
-✅ High — confirmed limitation, clear alternative
-
---
-
-## Dimension 4: GPS Update Rate
-
-### Fact Confirmation
-Per Fact #5, ArduPilot GPS_RATE_MS has a minimum of 200ms (5Hz). Our camera shoots at ~3fps (333ms). We produce a full VO+ESKF position estimate per frame at ~3Hz.
-
-### Reference Comparison
-3Hz < 5Hz minimum. ArduPilot's EKF expects at least 5Hz GPS updates for stable fusion.
-
-### Conclusion
-Between camera frames, use IMU prediction from the ESKF to interpolate position at 5Hz (or higher, e.g., 10Hz). The ESKF already runs IMU prediction at 100+Hz internally. Simply emit the ESKF predicted state as GPS_INPUT at 5-10Hz. Camera frame updates (3Hz) provide the measurement corrections. This is standard in sensor fusion: prediction runs fast, measurements arrive slower. The `fix_type` field can differentiate: camera-corrected frames → fix_type=3 (3D), IMU-predicted → fix_type=2 (2D) or adjust horiz_accuracy to reflect lower confidence.
-
-### Confidence
-✅ High — standard sensor fusion approach
-
---
-
-## Dimension 5: Startup & Failsafe
-
-### Fact Confirmation
-Per new AC: system initializes from last known GPS before GPS denial. On reboot: re-initialize from flight controller's IMU-extrapolated position. On total failure for N seconds: flight controller falls back to IMU-only.
-
-### Reference Comparison
-Draft02 has no startup or failsafe procedures. The system was assumed to already know its position at session start.
-
-### Conclusion
-Startup sequence:
-1. On boot, connect to flight controller via pymavlink
-2. Read current GPS position from flight controller (GLOBAL_POSITION_INT or GPS_RAW_INT message)
-3. Initialize ESKF state with this position
-4. Begin cuVSLAM initialization with first camera frames
-5. Start sending GPS_INPUT once ESKF has a valid position estimate
-
-Failsafe:
-1. If no position estimate for N seconds → stop sending GPS_INPUT (flight controller auto-detects GPS loss and falls back to IMU)
-2. Log failure event
-3. Continue attempting VO/satellite matching
-
-Reboot recovery:
-1. On companion computer reboot, reconnect to flight controller
-2. Read current GPS_RAW_INT (which is now IMU-extrapolated by flight controller)
-3. Re-initialize ESKF with this position (lower confidence)
-4. Resume normal operation
-
-### Confidence
-✅ High — standard autopilot integration patterns
-
---
-
-## Dimension 6: Object Localization
-
-### Fact Confirmation
-Per Fact #11, for oblique camera: ground_distance = altitude × tan(θ) where θ is angle from vertical. Combined with camera azimuth (yaw + camera pan angle) gives direction. With zoom, effective FOV narrows → higher pixel-to-meter resolution.
-
-### Reference Comparison
-Draft02 has basic nadir-only projection: pixel offset × GSD → meters → rotate by heading → lat/lon. The AI camera has configurable angle and zoom, so this needs extension.
-
-### Conclusion
-Object localization for AI camera:
-1. Get current UAV position from GPS-Denied system
-2. Get AI camera params: pan angle (azimuth relative to heading), tilt angle (from vertical), zoom level (→ effective focal length)
-3. Get pixel coordinates of detected object in AI camera frame
-4. Compute: a) bearing = UAV heading + camera pan angle + pixel horizontal offset angle, b) ground_distance = altitude / cos(tilt) × (tilt + pixel vertical offset angle) → simplified for flat terrain, c) convert bearing + distance to lat/lon offset from UAV position
-5. Accuracy inherits GPS-Denied position error + projection error from altitude/angle uncertainty
-
-Expose as lightweight local API (Unix socket or shared memory for speed, or simple HTTP on localhost).
-
-### Confidence
-✅ High — well-established trigonometry, flat terrain simplifies
-
---
-
-## Dimension 7: Thermal Management
-
-### Fact Confirmation
-Per Fact #9, Jetson Orin Nano Super throttles at 80°C junction temperature, dropping GPU from ~1GHz to ~300MHz (3x slowdown). Active cooling required. Power modes: 15W, 25W, MAXN SUPER.
-
-### Reference Comparison
-Draft02 ignores thermal constraints. Our pipeline (cuVSLAM ~9ms + satellite matcher ~50-200ms) runs on GPU continuously at 3fps. This is moderate but sustained load.
-
-### Conclusion
-Mitigation:
-1. Use 25W power mode (not MAXN SUPER) for stable sustained performance
-2. Require active cooling (5V fan, should be standard on any UAV companion computer mount)
-3. Monitor temperature via tegrastats/jtop at runtime
-4. If temp >75°C: reduce satellite matching frequency (every 5-10 frames instead of 3)
-5. If temp >80°C: skip satellite matching entirely, rely on VO+IMU only (cuVSLAM at 9ms is low power)
-6. Our total GPU time per 333ms frame: ~9ms cuVSLAM + ~50-200ms satellite match (async) = <60% GPU utilization → thermal throttling unlikely with proper cooling
-
-### Confidence
-⚠️ Medium — actual thermal behavior depends on airflow in UAV enclosure, ambient temperature in-flight
-
---
-
-## Dimension 8: VO Drift Budget Monitoring
-
-### Fact Confirmation
-New AC: max cumulative VO drift between satellite correction anchors < 100m. The ESKF maintains a position covariance matrix that grows during VO-only periods and shrinks on satellite corrections.
-
-### Reference Comparison
-Draft02 uses ESKF covariance for keyframe selection (trigger satellite match when covariance exceeds threshold) but doesn't explicitly track drift as a budget.
-
-### Conclusion
-Use ESKF position covariance diagonal (σ_x² + σ_y²) as the drift estimate. When √(σ_x² + σ_y²) approaches 100m:
-1. Force satellite matching on every frame (not just keyframes)
-2. Report LOW confidence via GPS_INPUT horiz_accuracy
-3. If drift exceeds 100m without satellite correction → flag as critical, increase matching frequency, send alert to ground station
-This is essentially what draft02 already does with covariance-based keyframe triggering, but now with an explicit 100m threshold.
-
-### Confidence
-✅ High — standard ESKF covariance interpretation
-
---
-
-## Dimension 9: Satellite Imagery Specs
-
-### Fact Confirmation
-New AC: ≥0.5 m/pixel resolution, <2 years old. Google Maps at zoom 18 = ~0.6 m/pixel, zoom 19 = ~0.3 m/pixel.
-
-### Reference Comparison
-Draft02 uses Google Maps zoom 18-19. Zoom 19 (0.3 m/pixel) exceeds the requirement. Zoom 18 (0.6 m/pixel) meets the minimum. Age depends on Google's imagery updates for eastern Ukraine — conflict zone may have stale imagery.
-
-### Conclusion
-Validate during offline preprocessing:
-1. Download at zoom 19 first (0.3 m/pixel)
-2. If zoom 19 unavailable for some tiles, fall back to zoom 18 (0.6 m/pixel — exceeds 0.5 minimum)
-3. Check imagery date metadata if available from Google Maps API
-4. Flag tiles where imagery appears stale (seasonal mismatch, destroyed buildings, etc.)
-5. No architectural change needed — add validation step to preprocessing pipeline
-
-### Confidence
-⚠️ Medium — Google Maps imagery age is not reliably queryable
-
---
-
-## Dimension 10: Internal IPC for Object Localization
-
-### Fact Confirmation
-Other onboard AI systems need to request GPS coordinates of detected objects. These systems run on the same Jetson.
-
-### Reference Comparison
-Draft02 has FastAPI for external API. For local IPC between processes on the same device, FastAPI is overkill but works.
-
-### Conclusion
-Retain a minimal FastAPI server on localhost:8000 for inter-process communication:
- POST /localize: accepts pixel coordinates + AI camera params → returns GPS coordinates
- GET /status: returns system health/state for monitoring
-This is local-only (bind to 127.0.0.1), not exposed externally. The primary output channel is MAVLink GPS_INPUT. This is a lightweight addition, not the core architecture.
-
-### Confidence
-✅ High — simple local IPC pattern
@@ -1,88 +0,0 @@
-# Validation Log
-
-## Validation Scenario
-A typical 15-minute flight over eastern Ukraine agricultural terrain. GPS is jammed after first 2 minutes. Flight includes straight segments, two sharp 90-degree turns, and one low-texture segment over a large plowed field. Ground station operator monitors via telemetry link. During the flight, companion computer reboots once due to power glitch.
-
-## Expected Based on Conclusions
-
-### Phase 1: Normal start (GPS available, first 2 min)
- System boots, connects to flight controller via pymavlink on UART
- Reads GLOBAL_POSITION_INT → initializes ESKF with real GPS position
- Begins cuVSLAM initialization with first camera frames
- Starts sending GPS_INPUT at 5Hz (ESKF prediction between frames)
- Ground station sees position + confidence via telemetry forwarding
-
-### Phase 2: GPS denial begins
- Flight controller's real GPS becomes unreliable/lost
- GPS-Denied system continues sending GPS_INPUT — seamless for autopilot
- horiz_accuracy changes from real-GPS level to VO-estimated level (~20m)
- cuVSLAM provides VO at every frame (~9ms), ESKF fuses with IMU
- Satellite matching runs every 3-10 frames on keyframes
- After successful satellite match: horiz_accuracy improves, fix_type stays 3
- NAMED_VALUE_FLOAT sends confidence/drift data to ground station at ~1Hz
-
-### Phase 3: Sharp turn
- cuVSLAM loses tracking (no overlapping features)
- ESKF falls back to IMU prediction, horiz_accuracy increases
- Next frame flagged as keyframe → satellite matching triggered immediately
- Satellite match against preloaded tiles using IMU dead-reckoning position
- If match found: position recovered, new segment begins, horiz_accuracy drops
- If 3 consecutive failures: send re-localization request to ground station via NAMED_VALUE_FLOAT/STATUSTEXT
- Ground station operator sends COMMAND_LONG with approximate coordinates
- System receives hint, constrains tile search → likely recovers position
-
-### Phase 4: Low-texture plowed field
- cuVSLAM keypoint count drops below threshold
- Satellite matching frequency increases (every frame)
- If satellite matching works on plowed field vs satellite imagery: position maintained
- If satellite also fails (seasonal difference): drift accumulates, ESKF covariance grows
- When √(σ²) approaches 100m: force continuous satellite matching
- horiz_accuracy reported as 50-100m, fix_type=2
-
-### Phase 5: Companion computer reboot
- Power glitch → Jetson reboots (~30-60 seconds)
- During reboot: flight controller gets no GPS_INPUT → detects GPS timeout → falls back to IMU-only dead reckoning
- Jetson comes back: reconnects via pymavlink, reads GPS_RAW_INT (IMU-extrapolated)
- Initializes ESKF with this position (low confidence, horiz_accuracy=100m)
- Begins cuVSLAM + satellite matching → gradually improves accuracy
- Operator on ground station sees position return with improving confidence
-
-### Phase 6: Object localization request
- AI detection system on same Jetson detects a vehicle in AI camera frame
- Sends POST /localize with pixel coords + camera angle (30° from vertical) + zoom level + altitude (500m)
- GPS-Denied system computes: ground_distance = 500 / cos(30°) = 577m slant, horizontal distance = 500 × tan(30°) = 289m
- Adds bearing from heading + camera pan → lat/lon offset
- Returns GPS coordinates with accuracy estimate (GPS-Denied accuracy + projection error)
-
-## Actual Validation Results
-The scenario covers all new AC requirements:
- ✅ MAVLink GPS_INPUT at 5Hz (camera frames + IMU interpolation)
- ✅ Confidence via horiz_accuracy field maps to confidence levels
- ✅ Ground station telemetry via MAVLink forwarding + NAMED_VALUE_FLOAT
- ✅ Re-localization via ground station command
- ✅ Startup from GPS → seamless transition on denial
- ✅ Reboot recovery from flight controller IMU-extrapolated position
- ✅ Drift budget tracking via ESKF covariance
- ✅ Object localization with AI camera angle/zoom
-
-## Counterexamples
-
-### Potential issue: 5Hz interpolation accuracy
-Between camera frames (333ms apart), ESKF predicts using IMU only. At 200km/h = 55m/s, the UAV moves ~18m between frames. IMU prediction over 200ms (one interpolation step) at this speed introduces ~1-5m error — acceptable for GPS_INPUT.
-
-### Potential issue: UART reliability
-Jetson Orin Nano UART instability reported (Fact #10). If MAVLink connection drops during flight, GPS_INPUT stops → autopilot loses GPS. Mitigation: use TCP over USB-C if UART unreliable, or add watchdog to reconnect. This is a hardware integration risk.
-
-### Potential issue: Telemetry bandwidth saturation
-If GPS-Denied sends too many NAMED_VALUE_FLOAT messages, it could compete with standard autopilot telemetry for bandwidth. Keep custom messages to 1Hz max (50-100 bytes/s = <1kbit/s).
-
-## Review Checklist
- [x] Draft conclusions consistent with fact cards
- [x] No important dimensions missed
- [x] No over-extrapolation
- [x] Conclusions actionable and verifiable
- [x] All new AC requirements addressed
- [ ] UART reliability needs hardware testing — cannot validate without physical setup
-
-## Conclusions Requiring Revision
-None — all conclusions hold under validation. The UART reliability risk needs flagging but doesn't change the architecture.
@@ -1,76 +0,0 @@
-# Acceptance Criteria Assessment
-
-## System Parameters (Calculated)
-
-| Parameter | Value |
-|-----------|-------|
-| GSD (at 400m) | 6.01 cm/pixel |
-| Ground footprint | 376m × 250m |
-| Consecutive overlap | 60-73% (at 100m intervals) |
-| Pixels per 50m | ~832 pixels |
-| Pixels per 20m | ~333 pixels |
-
-## Acceptance Criteria
-
-| Criterion | Our Values | Researched Values | Cost/Timeline Impact | Status |
-|-----------|-----------|-------------------|---------------------|--------|
-| GPS accuracy: 80% within 50m | 50m error for 80% of photos | NaviLoc: 19.5m MLE at 50-150m alt. Mateos-Ramirez: 143m mean at >1000m alt (with IMU). At 400m with 26MP + satellite correction, 50m for 80% is achievable with VO+SIM. No IMU adds ~30-50% error overhead. | Medium cost — needs robust satellite matching pipeline. ~3-4 weeks for core pipeline. | **Achievable** — keep as-is |
-| GPS accuracy: 60% within 20m | 20m error for 60% of photos | NaviLoc: 19.5m MLE at lower altitude (50-150m). At 400m, larger viewpoint gap increases error. Cross-view matching MA@20m improving +10% yearly. Needs high-quality satellite imagery and robust matching. | Higher cost — requires higher-quality satellite imagery (0.3-0.5m resolution). Additional 1-2 weeks for refinement. | **Challenging but achievable** — consider relaxing to 30m initially, tighten with iteration |
-| Handle 350m outlier photos | Tolerate up to 350m jump between consecutive photos | Standard VO systems detect outliers via feature matching failure. 350m at GSD 6cm = ~5833 pixels. Satellite re-localization can handle this if area is textured. | Low additional cost — outlier detection is standard in VO pipelines. | **Achievable** — keep as-is |
-| Sharp turns: <5% overlap, <200m drift, <70° angle | System continues working during sharp turns | <5% overlap means consecutive feature matching will fail. Must fall back to satellite matching for absolute position. At 400m altitude with 376m footprint, 200m drift means partial overlap with satellite. 70° rotation is large but manageable with rotation-invariant matchers (AKAZE, SuperPoint). | High complexity — requires multi-strategy architecture (VO primary, satellite fallback). +2-3 weeks. | **Achievable with architectural investment** — keep as-is |
-| Route disconnection & reconnection | Handle multiple disconnected route segments | Each segment needs independent satellite geo-referencing. Segments are stitched via common satellite reference frame. Similar to loop closure in SLAM but via external reference. | High complexity — core architectural challenge. +2-3 weeks for segment management. | **Achievable** — this should be a core design principle, not an edge case |
-| User input fallback (20% of route) | User provides GPS when system cannot determine | Simple UI interaction — user clicks approximate position on map. Becomes new anchor point. | Low cost — straightforward feature. | **Achievable** — keep as-is |
-| Processing speed: <5s per image | 5 seconds maximum per image | SuperPoint: ~50-100ms. LightGlue: ~20-50ms. Satellite crop+match: ~200-500ms. Full pipeline: ~500ms-2s on RTX 2060. NaviLoc runs 9 FPS on Raspberry Pi 5. ORB-SLAM3 with GPU: 30 FPS on Jetson TX2. | Low risk — well within budget on RTX 2060+. | **Easily achievable** — could target <2s. Keep 5s as safety margin |
-| Real-time streaming via SSE | Results appear immediately, refinement sent later | Standard architecture pattern. Process-and-stream is well-supported. | Low cost — standard web engineering. | **Achievable** — keep as-is |
-| Image Registration Rate > 95% | >95% of images successfully registered | ITU thesis: 93% SIM matching. With 60-73% consecutive overlap and deep learning features, >95% for VO between consecutive frames is achievable. The 5% tolerance covers sharp turns. | Medium cost — depends on feature matcher quality and satellite image quality. | **Achievable** — but interpret as "95% for normal consecutive frames". Sharp turn frames counted separately. |
-| MRE < 1.0 pixels | Mean Reprojection Error below 1 pixel | Sub-pixel accuracy is standard for SuperPoint/LightGlue. SVO achieves sub-pixel via direct methods. Typical range: 0.3-0.8 pixels. | No additional cost — inherent to modern matchers. | **Easily achievable** — keep as-is |
-| REST API + SSE background service | Always-running service, start on request, stream results | Standard Python (FastAPI) or .NET architecture. | Low cost — standard engineering. ~1 week for API layer. | **Achievable** — keep as-is |
-
-## Restrictions Assessment
-
-| Restriction | Our Values | Researched Values | Cost/Timeline Impact | Status |
-|-------------|-----------|-------------------|---------------------|--------|
-| No IMU data | No heading, no pitch/roll correction | **CRITICAL restriction.** Most published systems use IMU for heading and as fallback. Without IMU: (1) heading must be derived from consecutive frame matching or satellite matching, (2) no pitch/roll correction — rely on robust feature matchers, (3) scale from known altitude only. Adds ~30-50% error vs IMU-equipped systems. | High impact — requires visual heading estimation. All VO literature assumes at least heading from IMU. +2-3 weeks R&D for pure visual heading. | **Realistic but significantly harder.** Consider: can barometer data be available? |
-| Camera not auto-stabilized | Images have varying pitch/roll | At 400m with fixed-wing, typical roll ±15°, pitch ±10°. Causes trapezoidal distortion in images. Robust matchers (SuperPoint, LightGlue) handle moderate viewpoint changes. Homography estimation between frames compensates. | Medium impact — modern matchers handle this. Pre-rectification using estimated attitude could help. | **Realistic** — keep as-is. Mitigated by robust matchers. |
-| Google Maps only (cost-dependent) | Currently limited to Google Maps | Google Maps in eastern Ukraine may have 2-5 year old imagery. Conflict damage makes old imagery unreliable. **Risk: satellite-UAV matching may fail in areas with significant ground changes.** Alternatives: Mapbox (Maxar Vivid, sub-meter), Bing Maps (0.3-1m), Maxar SecureWatch (30cm, enterprise pricing). | High risk — may need multiple providers. Google: $200/month free credit. Mapbox: free tier for 100K requests. Maxar: enterprise pricing. | **Tighten** — add fallback provider. Pre-download tile cache for operational area. |
-| Image resolution FullHD to 6252×4168 | Variable resolution across flights | Lower resolution (FullHD=1920×1080) at 400m: GSD ≈ 0.20m/pixel, footprint ~384m × 216m. Significantly worse matching but still functional. Need to handle both extremes. | Medium impact — pipeline must be resolution-adaptive. | **Realistic** — keep. But note: FullHD accuracy will be ~3x worse than 26MP. |
-| Altitude ≤ 1km, terrain height negligible | Flat terrain assumption at known altitude | Simplifies scale estimation. At 400m, terrain variations of ±50m cause ±12.5% scale error. Eastern Ukraine is relatively flat (steppe), so this is reasonable. | Low impact for the operational area. | **Realistic** — keep as-is |
-| Mostly sunny weather | Good lighting conditions assumed | Sunny weather = good texture, consistent illumination. Shadows may cause matching issues but are manageable. | Low impact — favorable condition. | **Realistic** — keep. Add: "system performance degrades in overcast/low-light" |
-| Up to 3000 photos per flight | 500-1500 typical, 3000 maximum | At <5s per image: 3000 photos = ~4 hours max. Memory: 3000 × 26MP ≈ 78GB raw. Need efficient memory management and incremental processing. | Medium impact — requires streaming architecture and careful memory management. | **Realistic** — keep. Memory management is engineering, not research. |
-| Sharp turns with completely different next photo | Route discontinuity is possible | Most VO systems fail at 0% overlap. This is effectively a new "start point" problem. Satellite matching is the only recovery path. | High impact — already addressed in AC. | **Realistic** — this is the defining challenge |
-| Desktop/laptop with RTX 2060+ | Minimum GPU requirement | RTX 2060: 6GB VRAM, 1920 CUDA cores. Sufficient for SuperPoint, LightGlue, satellite matching. RTX 3070: 8GB VRAM, 5888 CUDA cores — significantly faster. | Low risk — hardware is adequate. | **Realistic** — keep as-is |
-
-## Missing Acceptance Criteria (Suggested Additions)
-
-| Criterion | Suggested Value | Rationale |
-|-----------|----------------|-----------|
-| Satellite imagery resolution requirement | ≥ 0.5 m/pixel, ideally 0.3 m/pixel | Matching quality depends heavily on reference imagery resolution. At GSD 6cm, satellite must be at least 0.5m for reliable cross-view matching. |
-| Confidence/uncertainty reporting | Report confidence score per position estimate | User needs to know which positions are reliable (satellite-anchored) vs uncertain (VO-only, accumulating drift). |
-| Output format | WGS84 coordinates in GeoJSON or CSV | Standardize output for downstream integration. |
-| Satellite image freshness requirement | < 2 years old for operational area | Older imagery may not match current ground truth due to conflict damage. |
-| Maximum drift between satellite corrections | < 100m cumulative VO drift before satellite re-anchor | Prevents long uncorrected VO segments from exceeding 50m target. |
-| Memory usage limit | < 16GB RAM, < 6GB VRAM | Ensures compatibility with RTX 2060 systems. |
-
-## Key Findings
-
-1. **The 50m/80% accuracy target is achievable** with a well-designed VO + satellite matching pipeline, even without IMU, given the high camera resolution (6cm GSD) and known altitude. NaviLoc achieves 19.5m at lower altitudes; our 400m altitude adds difficulty but 26MP resolution compensates.
-
-2. **The 20m/60% target is aggressive but possible** with high-quality satellite imagery (≤0.5m resolution). Consider starting with a 30m target and tightening through iteration. Performance heavily depends on satellite image quality and freshness for the operational area.
-
-3. **No IMU is the single biggest technical risk.** All published comparable systems use at least heading from IMU/magnetometer. Visual heading estimation from consecutive frames is feasible but adds noise. This restriction alone could require 2-3 extra weeks of R&D.
-
-4. **Google Maps satellite imagery for eastern Ukraine is a significant risk.** Imagery may be outdated (2-5 years) and may not reflect current ground conditions. A fallback satellite provider is strongly recommended.
-
-5. **Processing speed (<5s) is easily achievable** on RTX 2060+. Modern feature matching pipelines process in <500ms per pair. The pipeline could realistically achieve <2s per image.
-
-6. **Route disconnection handling should be the core architectural principle**, not an edge case. The system should be designed "segments-first" — each segment independently geo-referenced, then stitched.
-
-7. **Missing criterion: confidence reporting.** The user should see which positions are high-confidence (satellite-anchored) vs low-confidence (VO-extrapolated). This is critical for operational use.
-
-## Sources
- [Source #1] Mateos-Ramirez et al. (2024) — VO + satellite correction for fixed-wing UAV
- [Source #2] Öztürk (2025) — ORB-SLAM3 + SIM integration thesis
- [Source #3] NaviLoc (2025) — Trajectory-level visual localization
- [Source #4] LightGlue GitHub — Feature matching benchmarks
- [Source #5] DALGlue (2025) — Enhanced feature matching
- [Source #8-9] Satellite imagery coverage and pricing reports
@@ -1,63 +0,0 @@
-# Question Decomposition — AC & Restrictions Assessment
-
-## Original Question
-How realistic are the acceptance criteria and restrictions for a GPS-denied visual navigation system for fixed-wing UAV imagery?
-
-## Active Mode
-Mode A, Phase 1: AC & Restrictions Assessment
-
-## Question Type
-Knowledge Organization + Decision Support
-
-## Research Subject Boundary Definition
-
-| Dimension | Boundary |
-|-----------|----------|
-| **Platform** | Fixed-wing UAV, airplane type, not multirotor |
-| **Geography** | Eastern/southern Ukraine, left of Dnipro River (conflict zone, ~48.27°N, 37.38°E based on sample data) |
-| **Altitude** | ≤ 1km, sample data at 400m |
-| **Sensor** | Monocular RGB camera, 26MP, no IMU, no LiDAR |
-| **Processing** | Ground-based desktop/laptop with NVIDIA RTX 2060+ GPU |
-| **Time Window** | Current state-of-the-art (2024-2026) |
-
-## Problem Context Summary
-
-The system must determine GPS coordinates of consecutive aerial photo centers using only:
- Known starting GPS coordinates
- Known camera parameters (25mm focal, 23.5mm sensor, 6252×4168 resolution)
- Known flight altitude (≤1km, sample: 400m)
- Consecutive photos taken within ~100m of each other
- Satellite imagery (Google Maps) for ground reference
-
-Key constraints: NO IMU data, camera not auto-stabilized, potentially outdated satellite imagery for conflict zone.
-
-**Ground Sample Distance (GSD) at 400m altitude**:
- GSD = (400 × 23.5) / (25 × 6252) ≈ 0.060 m/pixel (6 cm/pixel)
- Ground footprint: ~376m × 250m per image
- Estimated consecutive overlap: 60-73% (depending on camera orientation relative to flight direction)
-
-## Sub-Questions for AC Assessment
-
-1. What GPS accuracy is achievable with VO + satellite matching at 400m altitude with 26MP camera?
-2. How does the absence of IMU affect accuracy and what compensations exist?
-3. What processing speed is achievable per image on RTX 2060+ for the required pipeline?
-4. What image registration rates are achievable with deep learning matchers?
-5. What reprojection errors are typical for modern feature matching?
-6. How do sharp turns and route disconnections affect VO systems?
-7. What satellite imagery quality is available for the operational area?
-8. What domain-specific acceptance criteria might be missing?
-
-## Timeliness Sensitivity Assessment
-
- **Research Topic**: GPS-denied visual navigation using deep learning feature matching
- **Sensitivity Level**: 🟠 High
- **Rationale**: Deep learning feature matchers (SuperPoint, LightGlue, GIM) are evolving rapidly; new methods appear quarterly. Satellite imagery providers update pricing and coverage frequently.
- **Source Time Window**: 12 months (2024-2026)
- **Priority official sources to consult**:
-  1. LightGlue GitHub repository (cvg/LightGlue)
-  2. ORB-SLAM3 documentation
-  3. Recent MDPI/IEEE papers on GPS-denied UAV navigation
- **Key version information to verify**:
-  - LightGlue: Current release and performance benchmarks
-  - SuperPoint: Compatibility and inference speed
-  - ORB-SLAM3: Monocular mode capabilities
@@ -1,133 +0,0 @@
-# Source Registry
-
-## Source #1
- **Title**: Visual Odometry in GPS-Denied Zones for Fixed-Wing UAV with Reduced Accumulative Error Based on Satellite Imagery
- **Link**: https://www.mdpi.com/2076-3417/14/16/7420
- **Tier**: L1
- **Publication Date**: 2024-08-22
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Fixed-wing UAV navigation researchers
- **Research Boundary Match**: ✅ Full match (fixed-wing, high altitude, satellite matching)
- **Summary**: VO + satellite image correction achieves 142.88m mean error over 17km at >1000m altitude using ORB + AKAZE. Uses IMU for heading and barometer for altitude. Error rate 0.83% of total distance.
- **Related Sub-question**: 1, 2
-
-## Source #2
- **Title**: Optimized visual odometry and satellite image matching-based localization for UAVs in GPS-denied environments (ITU Thesis)
- **Link**: https://polen.itu.edu.tr/items/1fe1e872-7cea-44d8-a8de-339e4587bee6
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV navigation researchers
- **Research Boundary Match**: ⚠️ Partial overlap (multirotor at 30-100m, but same VO+SIM methodology)
- **Summary**: ORB-SLAM3 + SuperPoint/SuperGlue/GIM achieves GPS-level accuracy. VO module: ±2m local accuracy. SIM module: 93% matching success rate. Demonstrated on DJI Mavic Air 2 at 30-100m.
- **Related Sub-question**: 1, 2, 4
-
-## Source #3
- **Title**: NaviLoc: Trajectory-Level Visual Localization for GNSS-Denied UAV Navigation
- **Link**: https://www.mdpi.com/2504-446X/10/2/97
- **Tier**: L1
- **Publication Date**: 2025-12
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV navigation / VPR researchers
- **Research Boundary Match**: ⚠️ Partial overlap (50-150m altitude, uses VIO not pure VO)
- **Summary**: Achieves 19.5m Mean Localization Error at 50-150m altitude. Runs at 9 FPS on Raspberry Pi 5. 16x improvement over AnyLoc-VLAD, 32x over raw VIO drift. Training-free system.
- **Related Sub-question**: 1, 7
-
-## Source #4
- **Title**: LightGlue: Local Feature Matching at Light Speed (GitHub + ICCV 2023)
- **Link**: https://github.com/cvg/LightGlue
- **Tier**: L1
- **Publication Date**: 2023 (actively maintained through 2025)
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Computer vision practitioners
- **Research Boundary Match**: ✅ Full match (core component)
- **Summary**: ~20-34ms per image pair on RTX 2080Ti. Adaptive pruning for fast inference. 2-4x speedup with PyTorch compilation.
- **Related Sub-question**: 3, 4
-
-## Source #5
- **Title**: Efficient image matching for UAV visual navigation via DALGlue
- **Link**: https://www.nature.com/articles/s41598-025-21602-5
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV navigation researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: DALGlue achieves 11.8% improvement over LightGlue on matching accuracy. Uses dual-tree complex wavelet preprocessing + linear attention for real-time performance.
- **Related Sub-question**: 3, 4
-
-## Source #6
- **Title**: Deep-UAV SLAM: SuperPoint and SuperGlue enhanced SLAM
- **Link**: https://isprs-archives.copernicus.org/articles/XLVIII-1-W5-2025/177/2025/
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV SLAM researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Replacing ORB-SLAM3's ORB features with SuperPoint+SuperGlue improved robustness and accuracy in aerial RGB scenarios.
- **Related Sub-question**: 4, 5
-
-## Source #7
- **Title**: SCAR: Satellite Imagery-Based Calibration for Aerial Recordings
- **Link**: https://arxiv.org/html/2602.16349v1
- **Tier**: L1
- **Publication Date**: 2026-02
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Aerial/satellite vision researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Long-term auto-calibration refinement by aligning aerial images with 2D-3D correspondences from orthophotos and elevation models.
- **Related Sub-question**: 1, 5
-
-## Source #8
- **Title**: Google Maps satellite imagery coverage and update frequency
- **Link**: https://ongeo-intelligence.com/blog/how-often-does-google-maps-update-satellite-images
- **Tier**: L3
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: GIS practitioners
- **Research Boundary Match**: ✅ Full match
- **Summary**: Conflict zones like eastern Ukraine face 2-5+ year update cycles. Imagery may be intentionally limited or blurred.
- **Related Sub-question**: 7
-
-## Source #9
- **Title**: Satellite Mapping Services comparison 2025
- **Link**: https://ts2.tech/en/exploring-the-world-from-above-top-satellite-mapping-services-for-web-mobile-in-2025/
- **Tier**: L3
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Developers, GIS practitioners
- **Research Boundary Match**: ✅ Full match
- **Summary**: Google: $200/month free credit, sub-meter resolution. Mapbox: Maxar imagery, generous free tier. Maxar SecureWatch: 30cm resolution, enterprise pricing. Planet: daily 3-4m imagery.
- **Related Sub-question**: 7
-
-## Source #10
- **Title**: Scale Estimation for Monocular Visual Odometry Using Reliable Camera Height
- **Link**: https://ieeexplore.ieee.org/document/9945178/
- **Tier**: L1
- **Publication Date**: 2022
- **Timeliness Status**: ✅ Currently valid (fundamental method)
- **Target Audience**: VO researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Known camera height/altitude resolves scale ambiguity in monocular VO. Essential for systems without IMU.
- **Related Sub-question**: 2
-
-## Source #11
- **Title**: Cross-View Geo-Localization benchmarks (SSPT, MA metrics)
- **Link**: https://www.mdpi.com/1424-8220/24/12/3719
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: VPR/geo-localization researchers
- **Research Boundary Match**: ⚠️ Partial overlap (general cross-view, not UAV-specific)
- **Summary**: SSPT achieved 84.40% RDS on UL14 dataset. MA improvements: +12% at 3m, +12% at 5m, +10% at 20m thresholds.
- **Related Sub-question**: 1
-
-## Source #12
- **Title**: ORB-SLAM3 GPU Acceleration Performance
- **Link**: https://arxiv.org/html/2509.10757v1
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: SLAM/VO engineers
- **Research Boundary Match**: ✅ Full match
- **Summary**: GPU acceleration achieves 2.8x speedup on desktop systems. 30 FPS achievable on Jetson TX2. Feature extraction up to 3x speedup with CUDA.
- **Related Sub-question**: 3
@@ -1,121 +0,0 @@
-# Fact Cards
-
-## Fact #1
- **Statement**: VO + satellite image correction achieves ~142.88m mean error over 17km flight at >1000m altitude using ORB features and AKAZE satellite matching. Error rate: 0.83% of total distance. This system uses IMU for heading and barometer for altitude.
- **Source**: Source #1 — https://www.mdpi.com/2076-3417/14/16/7420
- **Phase**: Phase 1
- **Target Audience**: Fixed-wing UAV at high altitude (>1000m)
- **Confidence**: ✅ High (peer-reviewed, real-world flight data)
- **Related Dimension**: GPS accuracy, drift correction
-
-## Fact #2
- **Statement**: ORB-SLAM3 monocular mode with optimized parameters achieves ±2m local accuracy for visual odometry. Scale ambiguity and drift remain for long flights.
- **Source**: Source #2 — ITU Thesis
- **Phase**: Phase 1
- **Target Audience**: UAV navigation (30-100m altitude, multirotor)
- **Confidence**: ✅ High (thesis with experimental validation)
- **Related Dimension**: VO accuracy, scale ambiguity
-
-## Fact #3
- **Statement**: Combined VO + Satellite Image Matching (SIM) with SuperPoint/SuperGlue/GIM achieves 93% matching success rate and "GPS-level accuracy" at 30-100m altitude.
- **Source**: Source #2 — ITU Thesis
- **Phase**: Phase 1
- **Target Audience**: Low-altitude UAV (30-100m)
- **Confidence**: ✅ High
- **Related Dimension**: Registration rate, satellite matching
-
-## Fact #4
- **Statement**: NaviLoc achieves 19.5m Mean Localization Error at 50-150m altitude, runs at 9 FPS on Raspberry Pi 5. 16x improvement over AnyLoc-VLAD. Training-free system.
- **Source**: Source #3 — NaviLoc paper
- **Phase**: Phase 1
- **Target Audience**: Low-altitude UAV (50-150m) in rural areas
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: GPS accuracy, processing speed
-
-## Fact #5
- **Statement**: LightGlue inference: ~20-34ms per image pair on RTX 2080Ti for 1024 keypoints. 2-4x speedup possible with PyTorch compilation and TensorRT.
- **Source**: Source #4 — LightGlue GitHub Issues
- **Phase**: Phase 1
- **Target Audience**: All GPU-accelerated vision systems
- **Confidence**: ✅ High (official repository benchmarks)
- **Related Dimension**: Processing speed
-
-## Fact #6
- **Statement**: SuperPoint+SuperGlue replacing ORB features in SLAM improves robustness and accuracy for aerial RGB imagery over classical handcrafted features.
- **Source**: Source #6 — ISPRS 2025
- **Phase**: Phase 1
- **Target Audience**: UAV SLAM researchers
- **Confidence**: ✅ High (peer-reviewed)
- **Related Dimension**: Feature matching quality
-
-## Fact #7
- **Statement**: Eastern Ukraine / conflict zones may have 2-5+ year old satellite imagery on Google Maps. Imagery may be intentionally limited, blurred, or restricted for security reasons.
- **Source**: Source #8
- **Phase**: Phase 1
- **Target Audience**: Ukraine conflict zone operations
- **Confidence**: ⚠️ Medium (general reporting, not Ukraine-specific verification)
- **Related Dimension**: Satellite imagery quality
-
-## Fact #8
- **Statement**: Maxar SecureWatch offers 30cm resolution with ~3M km² new imagery daily. Mapbox uses Maxar's Vivid imagery with sub-meter resolution. Google Maps offers sub-meter detail in urban areas but 1-3m in rural areas.
- **Source**: Source #9
- **Phase**: Phase 1
- **Target Audience**: All satellite imagery users
- **Confidence**: ✅ High
- **Related Dimension**: Satellite providers, cost
-
-## Fact #9
- **Statement**: Known camera height/altitude resolves scale ambiguity in monocular VO. The pixel-to-meter conversion is s = H / f × sensor_pixel_size, enabling metric reconstruction without IMU.
- **Source**: Source #10
- **Phase**: Phase 1
- **Target Audience**: Monocular VO systems
- **Confidence**: ✅ High (fundamental geometric relationship)
- **Related Dimension**: No-IMU compensation
-
-## Fact #10
- **Statement**: Camera heading (yaw) can be estimated from consecutive frame feature matching by decomposing the homography or essential matrix. Pitch/roll can be estimated from horizon detection or vanishing points. Without IMU, these estimates are noisier but functional.
- **Source**: Multiple vision-based heading estimation papers
- **Phase**: Phase 1
- **Target Audience**: Vision-only navigation systems
- **Confidence**: ⚠️ Medium (well-established but accuracy varies)
- **Related Dimension**: No-IMU compensation
-
-## Fact #11
- **Statement**: GSD at 400m with 25mm/23.5mm sensor/6252px = 6.01 cm/pixel. Ground footprint: 376m × 250m. At 100m photo interval, consecutive overlap is 60-73%.
- **Source**: Calculated from problem data using standard GSD formula
- **Phase**: Phase 1
- **Target Audience**: This specific system
- **Confidence**: ✅ High (deterministic calculation)
- **Related Dimension**: Image coverage, overlap
-
-## Fact #12
- **Statement**: GPU-accelerated ORB-SLAM3 achieves 2.8x speedup on desktop systems. 30 FPS possible on Jetson TX2. Feature extraction speedup up to 3x with CUDA-optimized pipelines.
- **Source**: Source #12
- **Phase**: Phase 1
- **Target Audience**: GPU-equipped systems
- **Confidence**: ✅ High
- **Related Dimension**: Processing speed
-
-## Fact #13
- **Statement**: Without IMU, the Mateos-Ramirez paper (Source #1) would lose: (a) yaw angle for rotation compensation, (b) fallback when feature matching fails. Their 142.88m error would likely be significantly higher without IMU heading data.
- **Source**: Inference from Source #1 methodology
- **Phase**: Phase 1
- **Target Audience**: This specific system
- **Confidence**: ⚠️ Medium (reasoned inference)
- **Related Dimension**: No-IMU impact
-
-## Fact #14
- **Statement**: DALGlue achieves 11.8% improvement over LightGlue on matching accuracy while maintaining real-time performance through dual-tree complex wavelet preprocessing and linear attention.
- **Source**: Source #5
- **Phase**: Phase 1
- **Target Audience**: Feature matching systems
- **Confidence**: ✅ High (peer-reviewed, 2025)
- **Related Dimension**: Feature matching quality
-
-## Fact #15
- **Statement**: Cross-view geo-localization benchmarks show MA@20m improving by +10% with latest methods (SSPT). RDS metric at 84.40% indicates reliable spatial positioning.
- **Source**: Source #11
- **Phase**: Phase 1
- **Target Audience**: Cross-view matching researchers
- **Confidence**: ✅ High
- **Related Dimension**: Cross-view matching accuracy
@@ -1,115 +0,0 @@
-# Comparison Framework
-
-## Selected Framework Type
-Decision Support (component-by-component solution comparison)
-
-## System Components
-1. Visual Odometry (consecutive frame matching)
-2. Satellite Image Geo-Referencing (cross-view matching)
-3. Heading & Orientation Estimation (without IMU)
-4. Drift Correction & Position Fusion
-5. Segment Management & Route Reconnection
-6. Interactive Point-to-GPS Lookup
-7. Pipeline Orchestration & API
-
---
-
-## Component 1: Visual Odometry
-
-| Solution | Tools | Advantages | Limitations | Fit |
-|----------|-------|-----------|-------------|-----|
-| ORB-SLAM3 monocular | ORB features, BA, map management | Mature, well-tested, handles loop closure. GPU-accelerated. 30FPS on Jetson TX2. | Scale ambiguity without IMU. Over-engineered for sequential aerial — map building not needed. Heavy dependency. | Medium — too complex for the use case |
-| Homography-based VO with SuperPoint+LightGlue | SuperPoint, LightGlue, OpenCV homography | Ground plane assumption perfect for flat terrain at 400m. Cleanly separates rotation/translation. Known altitude resolves scale directly. Fast. | Assumes planar scene (valid for our case). Fails at sharp turns (but that's expected). | **Best fit** — matches constraints exactly |
-| Optical flow VO | cv2.calcOpticalFlowPyrLK or RAFT | Dense motion field, no feature extraction needed. | Less accurate for large motions. Struggles with texture-sparse areas. No inherent rotation estimation. | Low — not suitable for 100m baselines |
-| Direct method (SVO) | SVO Pro | Sub-pixel precision, fast. | Designed for small baselines and forward cameras. Poor for downward aerial at large baselines. | Low |
-
-**Selected**: Homography-based VO with SuperPoint + LightGlue features
-
---
-
-## Component 2: Satellite Image Geo-Referencing
-
-| Solution | Tools | Advantages | Limitations | Fit |
-|----------|-------|-----------|-------------|-----|
-| SuperPoint + LightGlue cross-view matching | SuperPoint, LightGlue, perspective warp | Best overall performance on satellite stereo benchmarks. Fast (~50ms matching). Rotation-invariant. Handles viewpoint/scale changes. | Requires perspective warping to reduce viewpoint gap. Needs good satellite image quality. | **Best fit** — proven on satellite imagery |
-| SuperPoint + SuperGlue + GIM | SuperPoint, SuperGlue, GIM | GIM adds generalization for challenging scenes. 93% match rate (ITU thesis). | SuperGlue slower than LightGlue. GIM adds complexity. | Good — slightly better robustness, slower |
-| LoFTR (detector-free) | LoFTR | No keypoint detection step. Works on low-texture. | Slower than detector-based methods. Fixed resolution (coarse). Less accurate than SuperPoint+LightGlue on satellite benchmarks. | Medium — fallback option |
-| DUSt3R/MASt3R | DUSt3R/MASt3R | Handles extreme viewpoints and low overlap. +50% completeness over COLMAP in sparse scenarios. | Very slow. Designed for 3D reconstruction not 2D matching. Unreliable with many images. | Low — only for extreme fallback |
-| Terrain-weighted optimization (YFS90) | Custom pipeline + DEM | <7m MAE without IMU! Drift-free. Handles thermal IR. 20 scenarios validated. | Requires DEM data. More complex implementation. Not open-source matching details. | High — architecture inspiration |
-
-**Selected**: SuperPoint + LightGlue (primary) with perspective warping. GIM as supplementary for difficult matches. YFS90-style terrain-weighted sliding window for position optimization.
-
---
-
-## Component 3: Heading & Orientation Estimation
-
-| Solution | Tools | Advantages | Limitations | Fit |
-|----------|-------|-----------|-------------|-----|
-| Homography decomposition (consecutive frames) | OpenCV decomposeHomographyMat | Directly gives rotation between frames. Works with ground plane assumption. No extra sensors needed. | Accumulates heading drift over time. Noisy for small motions. Ambiguous decomposition (need to select correct solution). | **Best fit** — primary heading source |
-| Satellite matching absolute orientation | From satellite match homography | Provides absolute heading correction. Eliminates accumulated heading drift. | Only available when satellite match succeeds. Intermittent. | **Best fit** — drift correction for heading |
-| Optical flow direction | Dense flow vectors | Simple to compute. | Very noisy at high altitude. Unreliable for heading. | Low |
-
-**Selected**: Homography decomposition for frame-to-frame heading + satellite matching for periodic absolute heading correction.
-
---
-
-## Component 4: Drift Correction & Position Fusion
-
-| Solution | Tools | Advantages | Limitations | Fit |
-|----------|-------|-----------|-------------|-----|
-| Kalman filter (EKF/UKF) | filterpy or custom | Well-understood. Handles noisy measurements. Good for fusing VO + satellite. | Assumes Gaussian noise. Linearization issues with EKF. | Good — simple and effective |
-| Sliding window optimization with terrain constraints | Custom optimization, scipy.optimize | YFS90 achieves <7m with this. Directly constrains drift. No loop closure needed. | More complex to implement. Needs tuning. | **Best fit** — proven for this exact problem |
-| Pose graph optimization | g2o, GTSAM | Standard in SLAM. Handles satellite anchors as prior factors. Globally optimal. | Heavy dependency. Over-engineered if segments are short. | Medium — overkill unless routes are very long |
-| Simple anchor reset | Direct correction at satellite match | Simplest. Just replace VO position with satellite position. | Discontinuous trajectory. No smoothing. | Low — too crude |
-
-**Selected**: Sliding window optimization with terrain constraints (inspired by YFS90), with Kalman filter as simpler fallback. Satellite matches as absolute anchor constraints.
-
---
-
-## Component 5: Segment Management & Route Reconnection
-
-| Solution | Tools | Advantages | Limitations | Fit |
-|----------|-------|-----------|-------------|-----|
-| Segments-first architecture with satellite anchoring | Custom segment manager | Each segment independently geo-referenced. No dependency between disconnected segments. Natural handling of sharp turns. | Needs robust satellite matching per segment. Segments without any satellite match are "floating". | **Best fit** — matches AC requirement for core strategy |
-| Global pose graph with loop closure | g2o/GTSAM | Can connect segments when they revisit same area. | Heavy. Doesn't help if segments don't overlap with each other. | Low — segments may not revisit same areas |
-| Trajectory-level VPR (NaviLoc-style) | VPR + trajectory optimization | Global optimization across trajectory. | Requires pre-computed VPR database. Complex. Designed for continuous trajectory, not disconnected segments. | Low |
-
-**Selected**: Segments-first architecture. Each segment starts from a satellite anchor or user input. Segments connected through shared satellite coordinate frame.
-
---
-
-## Component 6: Interactive Point-to-GPS Lookup
-
-| Solution | Tools | Advantages | Limitations | Fit |
-|----------|-------|-----------|-------------|-----|
-| Homography projection (image → ground) | Computed homography from satellite match | Already computed during geo-referencing. Accurate for flat terrain. | Only works for images with successful satellite match. | **Best fit** |
-| Camera ray-casting with known altitude | Camera intrinsics + pose estimate | Works for any image with pose estimate. Simpler math. | Accuracy depends on pose estimate quality. | Good — fallback for non-satellite-matched images |
-
-**Selected**: Homography projection (primary) + ray-casting (fallback).
-
---
-
-## Component 7: Pipeline & API
-
-| Solution | Tools | Advantages | Limitations | Fit |
-|----------|-------|-----------|-------------|-----|
-| Python FastAPI + SSE | FastAPI, EventSourceResponse, asyncio | Native SSE support (since 0.135.0). Async GPU pipeline. Excellent for ML/CV workloads. Rich ecosystem. | Python GIL (mitigated with async/multiprocessing). | **Best fit** — natural for CV/ML pipeline |
-| .NET ASP.NET Core + SSE | ASP.NET Core, SignalR | High performance. Good for enterprise. | Less natural for CV/ML. Python interop needed for PyTorch models. Adds complexity. | Low — unnecessary indirection |
-| Python + gRPC streaming | gRPC | Efficient binary protocol. Bidirectional streaming. | More complex client integration. No browser-native support. | Medium — overkill for this use case |
-
-**Selected**: Python FastAPI with SSE.
-
---
-
-## Google Maps Tile Resolution at Latitude 48° (Operational Area)
-
-| Zoom Level | Meters/pixel | Tile coverage (256px) | Tiles for 20km² | Download size est. |
-|-----------|-------------|----------------------|-----------------|-------------------|
-| 17 | 0.80 m/px | ~205m × 205m | ~500 tiles | ~20MB |
-| 18 | 0.40 m/px | ~102m × 102m | ~2,000 tiles | ~80MB |
-| 19 | 0.20 m/px | ~51m × 51m | ~8,000 tiles | ~320MB |
-| 20 | 0.10 m/px | ~26m × 26m | ~30,000 tiles | ~1.2GB |
-
-Formula: metersPerPx = 156543.03 × cos(48° × π/180) / 2^zoom ≈ 104,771 / 2^zoom
-
-**Selected**: Zoom 18 (0.40 m/px) as primary matching resolution. Zoom 19 (0.20 m/px) for refinement if available. Meets the ≥0.5 m/pixel AC requirement.
@@ -1,146 +0,0 @@
-# Reasoning Chain
-
-## Dimension 1: GPS Accuracy (50m/80%, 20m/60%)
-
-### Fact Confirmation
- YFS90 system achieves <7m MAE without IMU (Fact from Source DOAJ/GitHub)
- NaviLoc achieves 19.5m MLE at 50-150m altitude (Fact #4)
- Mateos-Ramirez achieves 143m mean error at >1000m altitude with IMU (Fact #1)
- Our GSD is 6cm/pixel at 400m altitude (Fact #11)
- ITU thesis achieves GPS-level accuracy with VO+SIM at 30-100m (Fact #3)
-
-### Reference Comparison
- At 400m altitude, our camera produces much higher resolution imagery than typical systems
- YFS90 at <7m without IMU is the strongest reference — uses terrain-weighted constraint optimization
- NaviLoc at 19.5m uses trajectory-level optimization but at lower altitude
- The combination of VO + satellite matching with sliding window optimization should achieve 10-30m depending on satellite image quality
-
-### Conclusion
- **50m / 80%**: High confidence achievable. Multiple systems achieve better than this.
- **20m / 60%**: Achievable with good satellite imagery. YFS90 achieves <7m. Our higher altitude makes cross-view matching harder, but 26MP camera compensates.
- **10m stretch**: Possible with zoom 19 satellite tiles (0.2m/px) and terrain-weighted optimization.
-
-### Confidence: ✅ High for 50m, ⚠️ Medium for 20m, ❓ Low for 10m
-
---
-
-## Dimension 2: No-IMU Heading Estimation
-
-### Fact Confirmation
- Homography decomposition gives rotation between frames for planar scenes (multiple sources)
- Ground plane assumption is valid for flat terrain (eastern Ukraine steppe)
- Satellite matching provides absolute orientation correction (Sources #1, #2)
- YFS90 achieves <7m without requiring IMU (Source #3 DOAJ)
-
-### Reference Comparison
- Most published systems use IMU for heading — our approach is less common
- YFS90 proves it's possible without IMU, but uses DEM data for terrain weighting
- The key insight: satellite matching provides both position AND heading correction, making intermittent heading drift from VO acceptable
-
-### Conclusion
-Heading estimation from homography decomposition between consecutive frames + periodic satellite matching correction is viable. The frame-to-frame heading drift accumulates, but satellite corrections at regular intervals (every 5-20 frames) reset it. The flat terrain of the operational area makes the ground plane assumption reliable.
-
-### Confidence: ⚠️ Medium — novel approach but supported by YFS90 results
-
---
-
-## Dimension 3: Processing Speed (<5s per image)
-
-### Fact Confirmation
- LightGlue: ~20-50ms per pair (Fact #5)
- SuperPoint extraction: ~50-100ms per image
- GPU-accelerated ORB-SLAM3: 30 FPS (Fact #12)
- NaviLoc: 9 FPS on Raspberry Pi 5 (Fact #4)
-
-### Pipeline Time Budget Estimate (per image on RTX 2060)
-1. SuperPoint feature extraction: ~80ms
-2. LightGlue VO matching (vs previous frame): ~40ms
-3. Homography estimation + position update: ~5ms
-4. Satellite tile crop (from cache): ~10ms
-5. SuperPoint extraction on satellite crop: ~80ms
-6. LightGlue satellite matching: ~60ms
-7. Position correction + sliding window optimization: ~20ms
-8. Total: ~295ms ≈ 0.3s
-
-### Conclusion
-Processing comfortably fits within 5s budget. Even with additional overhead (satellite tile download, perspective warping, GIM fallback), the pipeline stays under 2s. The 5s budget provides ample margin.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 4: Sharp Turns & Route Disconnection
-
-### Fact Confirmation
- At <5% overlap, consecutive feature matching will fail
- Satellite matching can provide absolute position independently of VO
- DUSt3R/MASt3R handle extreme low overlap (+50% completeness vs COLMAP)
- YFS90 handles positioning failures with re-localization
-
-### Reference Comparison
- Traditional VO systems fail at sharp turns — this is expected and acceptable
- The segments-first architecture treats each continuous VO chain as a segment
- Satellite matching re-localizes at the start of each new segment
- If satellite matching fails too → wider search area → user input
-
-### Conclusion
-The system should not try to match across sharp turns. Instead:
-1. Detect VO failure (low match count / high reprojection error)
-2. Start new segment
-3. Attempt satellite geo-referencing for new segment start
-4. Each segment is independently positioned in the global satellite coordinate frame
-
-This is architecturally simpler and more robust than trying to bridge disconnections.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 5: Satellite Image Matching Reliability
-
-### Fact Confirmation
- Google Maps at zoom 18: 0.40 m/px at lat 48° — meets AC requirement
- Eastern Ukraine imagery may be 2-5 years old (Fact #7)
- SuperPoint+LightGlue is best performer for satellite matching (Source comparison study)
- Perspective warping improves cross-view matching significantly
- 93% match rate achieved in ITU thesis (Fact #3)
-
-### Reference Comparison
- The main risk is satellite image freshness in conflict zone
- Natural terrain features (rivers, forests, field boundaries) are relatively stable over years
- Man-made features (buildings, roads) may change due to conflict
- Agricultural field patterns change seasonally
-
-### Conclusion
-Satellite matching will work reliably in areas with stable natural features. Performance degrades in:
-1. Areas with significant conflict damage (buildings destroyed)
-2. Areas with seasonal agricultural changes
-3. Areas with very homogeneous texture (large uniform fields)
-
-Mitigation: use multiple scale levels, widen search area, accept lower confidence.
-
-### Confidence: ⚠️ Medium — depends heavily on operational area characteristics
-
---
-
-## Dimension 6: Architecture Selection
-
-### Fact Confirmation
- YFS90 architecture (VO + satellite matching + terrain-weighted optimization) achieves <7m
- ITU thesis architecture (ORB-SLAM3 + SIM) achieves GPS-level accuracy
- NaviLoc architecture (VPR + trajectory optimization) achieves 19.5m
-
-### Reference Comparison
- YFS90 is closest to our requirements: no IMU, satellite matching, drift correction
- Our system adds: segment management, real-time streaming, user fallback
- We need simpler VO than ORB-SLAM3 (no map building needed)
- We need faster matching than SuperGlue (LightGlue preferred)
-
-### Conclusion
-Hybrid architecture combining:
- YFS90-style sliding window optimization for drift correction
- SuperPoint + LightGlue for both VO and satellite matching (unified feature pipeline)
- Segments-first architecture for disconnection handling
- FastAPI + SSE for real-time streaming
-
-### Confidence: ✅ High
@@ -1,57 +0,0 @@
-# Validation Log
-
-## Validation Scenario
-Using the provided sample data: 60 consecutive images from a flight starting at (48.275292, 37.385220) heading generally south-southwest. Camera: 26MP at 400m altitude.
-
-## Expected Behavior Based on Conclusions
-
-### Normal consecutive frames (AD000001-AD000032)
- VO successfully matches consecutive frames (60-73% overlap)
- Satellite matching every 5-10 frames provides absolute correction
- Position error stays within 20-50m corridor around ground truth
- Heading estimated from homography, corrected by satellite matching
-
-### Apparent maneuver zone (AD000033-AD000048)
- The coordinates show the UAV making a complex turn around images 33-48
- Some consecutive pairs may have low overlap → VO quality drops
- Satellite matching becomes the primary position source
- New segments may be created if VO fails completely
- Position confidence drops in this zone
-
-### Return to straight flight (AD000049-AD000060)
- VO re-establishes strong consecutive matching
- Satellite matching re-anchors position
- Accuracy returns to normal levels
-
-## Actual Validation (Calculated)
-
-Distances between consecutive samples in the data:
- AD000001→002: ~180m (larger than stated 100m — likely exaggeration in problem description)
- AD000002→003: ~115m
- Typical gap: 80-180m
- At 376m footprint width and 250m height, even 180m gap gives 52-73% overlap → sufficient for VO
-
-At the turn zone (images 33-48):
- AD000041→042: ~230m with direction change → overlap may drop to 30-40%
- AD000042→043: ~230m with direction change → overlap may drop significantly
- AD000045→046: ~160m with direction change → may be <20% overlap
- These transitions are where VO may fail → satellite matching needed
-
-## Counterexamples
-
-1. **Homogeneous terrain**: If a section of the flight is over large uniform agricultural fields with no distinguishing features, both VO and satellite matching may fail. Mitigation: use higher zoom satellite tiles, rely on VO with lower confidence.
-
-2. **Conflict-damaged area**: If satellite imagery shows pre-war structures that no longer exist, satellite matching will produce incorrect position estimates. Mitigation: confidence scoring will flag inconsistent matches.
-
-3. **FullHD resolution flight**: At GSD 20cm/pixel instead of 6cm, matching quality degrades ~3x. The 50m target may still be achievable but 20m will be very difficult.
-
-## Review Checklist
- [x] Draft conclusions consistent with fact cards
- [x] No important dimensions missed
- [x] No over-extrapolation
- [x] Issue found: The problem states "within 100 meters of each other" but actual data shows 80-230m. Pipeline must handle larger baselines.
- [x] Issue found: Tile download strategy needs to handle unknown route direction — progressive expansion needed.
-
-## Conclusions Requiring Revision
- Photo spacing is 80-230m not strictly 100m — increases the range of overlap variations. Still functional but wider variance than assumed.
- Route direction is unknown at start — satellite tile pre-loading must use expanding radius strategy, not directional pre-loading.
@@ -1,73 +0,0 @@
-# Question Decomposition
-
-## Original Question
-"Analyze completeness of the current solution. How mature is it?"
-
-## Active Mode
-**Mode B: Solution Assessment** — 5 solution drafts exist (`solution_draft01.md` through `solution_draft05.md`). Assessing the latest draft (05) for completeness and maturity.
-
-## Question Type Classification
-**Knowledge Organization + Problem Diagnosis**
- Knowledge Organization: systematically map what a complete GPS-denied nav system requires vs what is present
- Problem Diagnosis: identify gaps, weak points, and missing elements that reduce maturity
-
-## Research Subject Boundary Definition
-
-| Dimension | Boundary |
-|-----------|----------|
-| **Population** | Fixed-wing UAV GPS-denied visual navigation systems using visual odometry + satellite matching + IMU fusion |
-| **Geography** | Eastern/southern Ukraine conflict zone operations |
-| **Timeframe** | Current state of art (2024-2026), focusing on Jetson-class embedded deployment |
-| **Level** | System-level architecture completeness — from sensor input to flight controller output |
-
-## Problem Context Summary
-
-The solution is a real-time GPS-denied visual navigation system for a custom 3.5m fixed-wing UAV:
- **Hardware**: Jetson Orin Nano Super (8GB), ADTI 20L V1 camera (0.7fps), Viewpro A40 Pro gimbal, Pixhawk 6x
- **Core pipeline**: cuVSLAM VO (0.7fps) → ESKF fusion → GPS_INPUT via pymavlink at 5-10Hz
- **Satellite correction**: LiteSAM/EfficientLoFTR/XFeat TRT FP16 on keyframes, async Stream B
- **5 draft iterations**: progressed from initial architecture → TRT migration → camera rate correction + UAV platform specs
- **Supporting docs**: tech_stack.md, security_analysis.md
-
-## Decomposed Sub-Questions (Mode B)
-
-### Functional Completeness
- **SQ-1**: What components does a mature GPS-denied visual navigation system require that are missing or under-specified in the current draft?
- **SQ-2**: How complete is the ESKF sensor fusion specification? (state vector, process model, measurement models, Q/R tuning, observability analysis)
- **SQ-3**: How does the system handle disconnected route segments (sharp turns with no overlap)? Is this adequately specified?
- **SQ-4**: What coordinate system transformations are needed (camera → body → NED → WGS84) and are they specified?
- **SQ-5**: How does the system handle initial localization (first frame + satellite matching bootstrap)?
- **SQ-6**: Is the re-localization request workflow (to ground station) sufficiently defined?
- **SQ-7**: How complete is the offline tile preparation pipeline (zoom levels, storage requirements, coverage calculation)?
- **SQ-8**: Is the object localization component sufficiently specified for operational use?
-
-### Performance & Robustness
- **SQ-9**: What are the realistic drift characteristics of cuVSLAM at 0.7fps over long straight segments?
- **SQ-10**: How robust is satellite matching with Google Maps imagery in the operational area?
- **SQ-11**: What happens during extended periods with no satellite match (cloud cover on tiles, homogeneous terrain)?
- **SQ-12**: Is the 5-10Hz GPS_INPUT rate adequate for the flight controller's EKF?
-
-### Maturity Assessment
- **SQ-13**: What is the Technology Readiness Level (TRL) of each component?
- **SQ-14**: What validation/testing has been done vs what is only planned?
- **SQ-15**: What operational procedures are missing (pre-flight checklist, in-flight monitoring, post-flight analysis)?
- **SQ-16**: Are there any inconsistencies between documents (tech_stack.md, security_analysis.md, solution_draft05.md)?
-
-### Security
- **SQ-17**: Are there security gaps not covered by the existing security_analysis.md?
- **SQ-18**: How does the MAVLink GPS_INPUT message security work (spoofing of the GPS replacement itself)?
-
-## Timeliness Sensitivity Assessment
-
- **Research Topic**: GPS-denied visual navigation system completeness and maturity
- **Sensitivity Level**: 🟡 Medium
- **Rationale**: Core algorithms (VO, ESKF, feature matching) are well-established. Hardware (Jetson Orin Nano Super) is relatively new but stable. cuVSLAM library updates are moderate pace. No rapidly-changing AI/LLM dependencies.
- **Source Time Window**: 1-2 years
- **Priority official sources to consult**:
-  1. NVIDIA cuVSLAM / Isaac ROS documentation
-  2. PX4/ArduPilot MAVLink GPS_INPUT documentation
-  3. LiteSAM / EfficientLoFTR papers and repos
- **Key version information to verify**:
-  - cuVSLAM: PyCuVSLAM v15.0.0
-  - TensorRT: 10.3.0
-  - JetPack: 6.2.2
@@ -1,166 +0,0 @@
-# Source Registry
-
-## Source #1
- **Title**: ArduPilot AP_GPS_Params — GPS_RATE minimum 5Hz
- **Link**: https://github.com/ArduPilot/ardupilot/pull/15980
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: ArduPilot flight controller users
- **Research Boundary Match**: ✅ Full match
- **Summary**: ArduPilot enforces minimum 5Hz GPS update rate. GPS_RATE parameter description: "Lowering below 5Hz(default) is not allowed."
- **Related Sub-question**: SQ-12
-
-## Source #2
- **Title**: MAVLink GPS_INPUT Message Definition
- **Link**: https://ardupilot.org/mavproxy/docs/modules/GPSInput.html
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: MAVLink developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: GPS_INPUT requires: lat, lon, alt, fix_type, hdop, vdop, horiz_accuracy, vert_accuracy, speed_accuracy, vn, ve, vd, time_usec, time_week, time_week_ms, satellites_visible, gps_id, ignore_flags. GPS_TYPE=14 for MAVLink GPS.
- **Related Sub-question**: SQ-6, SQ-12
-
-## Source #3
- **Title**: pymavlink GPS_INPUT example (GPS_INPUT_pymavlink.py)
- **Link**: https://webperso.ensta.fr/lebars/Share/GPS_INPUT_pymavlink.py
- **Tier**: L3
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: pymavlink developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Working pymavlink example sending GPS_INPUT over serial at 10Hz with GPS time calculation from system time.
- **Related Sub-question**: SQ-6
-
-## Source #4
- **Title**: PyCuVSLAM API Reference (v15.0.0)
- **Link**: https://wiki.seeedstudio.com/pycuvslam_recomputer_robotics/
- **Tier**: L2
- **Publication Date**: 2026-03
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: cuVSLAM developers on Jetson
- **Research Boundary Match**: ✅ Full match
- **Summary**: cuVSLAM supports mono/stereo/inertial modes. Requires Camera model (fx,fy,cx,cy,distortion), ImuCalibration (noise density, random walk, frequency, T_imu_rig). Modes: Performance/Precision/Moderate. IMU fallback ~1s acceptable quality.
- **Related Sub-question**: SQ-1, SQ-5, SQ-9
-
-## Source #5
- **Title**: ESKF Python implementation for fixed-wing UAV
- **Link**: https://github.com/ludvigls/ESKF
- **Tier**: L4
- **Publication Date**: 2023
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: ESKF implementers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Reference ESKF: 16-state vector (pos[3], vel[3], quat[4], acc_bias[3], gyro_bias[3]). Prediction with IMU at high rate. Update with GPS position/velocity. Tuning parameters: Q (process noise), R (measurement noise).
- **Related Sub-question**: SQ-2
-
-## Source #6
- **Title**: ROS ESKF based on PX4/ecl — multi-sensor fusion
- **Link**: https://github.com/EliaTarasov/ESKF
- **Tier**: L4
- **Publication Date**: 2022
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV ESKF implementers
- **Research Boundary Match**: ✅ Full match
- **Summary**: ESKF fusing GPS, Magnetometer, Vision Pose, Optical Flow, RangeFinder with IMU. Shows that vision pose and optical flow are separate measurement models, each with its own observation matrix and noise parameters.
- **Related Sub-question**: SQ-2
-
-## Source #7
- **Title**: Visual-Inertial Odometry Scale Observability (Range-VIO)
- **Link**: https://arxiv.org/abs/2103.15215
- **Tier**: L1
- **Publication Date**: 2021
- **Timeliness Status**: ✅ Currently valid (fundamental research)
- **Target Audience**: VIO researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Monocular VIO cannot observe metric scale without accelerometer excitation (not constant velocity). A 1D range sensor makes scale observable. For our case, barometric altitude + known flight altitude provides this constraint.
- **Related Sub-question**: SQ-2, SQ-4
-
-## Source #8
- **Title**: NaviLoc: Trajectory-Level Visual Localization for GNSS-Denied UAVs
- **Link**: https://www.mdpi.com/2504-446X/10/2/97
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: GPS-denied UAV navigation researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Trajectory-level optimization fusing VPR with VIO achieves 19.5m mean error at 50-150m altitude. Key insight: treating satellite matching as noisy measurement rather than ground truth, with trajectory-level optimization. Runs at 9 FPS on RPi 5.
- **Related Sub-question**: SQ-3, SQ-13
-
-## Source #9
- **Title**: SatLoc-Fusion: Hierarchical Adaptive Fusion Framework
- **Link**: https://www.scilit.com/publications/e5cafaf875a49297a62b298a89d5572f
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: GPS-denied UAV researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Three-layer fusion: absolute geo-localization (DinoV2), relative VO (XFeat), optical flow velocity. Adaptive weighting based on confidence. Achieves <15m error, >90% trajectory coverage. 2Hz on 6 TFLOPS edge.
- **Related Sub-question**: SQ-3, SQ-10, SQ-13
-
-## Source #10
- **Title**: Auterion GPS-Denied Workflow
- **Link**: https://docs.auterion.com/vehicle-operation/auterion-mission-control/useful-resources/operations/gps-denied-workflow
- **Tier**: L2
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: UAV operators
- **Research Boundary Match**: ⚠️ Partial overlap (multirotor focus, but procedures applicable)
- **Summary**: Pre-flight: manually set home position, reset heading/position, configure wind. In-flight: enable INS mode. Defines operational procedures for GPS-denied missions.
- **Related Sub-question**: SQ-15
-
-## Source #11
- **Title**: PX4 GNSS-Degraded & Denied Flight (Dead-Reckoning)
- **Link**: https://docs.px4.io/main/en/advanced_config/gnss_degraded_or_denied_flight.html
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: PX4 users
- **Research Boundary Match**: ⚠️ Partial overlap (PX4-specific, but concepts apply to ArduPilot)
- **Summary**: GPS-denied requires redundant position/velocity sensors. Dead-reckoning mode for intermittent GNSS loss. Defines failsafe behaviors when GPS is lost.
- **Related Sub-question**: SQ-15
-
-## Source #12
- **Title**: Google Maps Ukraine satellite imagery coverage
- **Link**: https://newsukraine.rbc.ua/news/google-maps-has-surprise-for-satellite-imagery-1727182380.html
- **Tier**: L3
- **Publication Date**: 2024-09
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: General public
- **Research Boundary Match**: ✅ Full match
- **Summary**: Google Maps improved imagery quality with Cloud Score+ AI. However, conflict zone imagery is intentionally older (>1 year). Ukrainian officials flagged security concerns about imagery revealing military positions.
- **Related Sub-question**: SQ-10
-
-## Source #13
- **Title**: Jetson Orin Nano Super thermal behavior at 25W
- **Link**: https://edgeaistack.app/blog/jetson-orin-nano-power-consumption/
- **Tier**: L3
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Jetson developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Thermal throttling at SoC junction >80°C. Sustained GPU at 25W: ~50-51°C reported. Active cooling required for >15W. Most production workloads 8-15W.
- **Related Sub-question**: SQ-11
-
-## Source #14
- **Title**: Automated Image Matching for Satellite Images with Different GSDs
- **Link**: https://www.kjrs.org/journal/view.html?pn=related&uid=756&vmd=Full
- **Tier**: L1
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Remote sensing researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: GSD mismatch between satellite and aerial images requires scale normalization via subsampling/super-resolution. Coarse-to-fine matching strategy effective. Scale-invariant features (SIFT, deep features) partially handle scale differences.
- **Related Sub-question**: SQ-7
-
-## Source #15
- **Title**: Optimized VO and satellite image matching for UAVs (Istanbul Tech thesis)
- **Link**: https://polen.itu.edu.tr/items/1fe1e872-7cea-44d8-a8de-339e4587bee6
- **Tier**: L1
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: GPS-denied UAV researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Complete VO+satellite matching pipeline. Coordinate transforms: GPS → local NED for trajectory comparison. PnP solver for UAV pose from correspondences. Map retrieval using VO-estimated position to crop satellite tiles.
- **Related Sub-question**: SQ-4, SQ-5
@@ -1,169 +0,0 @@
-# Fact Cards
-
-## Fact #1 — ArduPilot minimum GPS rate is 5Hz
- **Statement**: ArduPilot enforces a hard minimum of 5Hz for GPS_INPUT updates. The GPS_RATE parameter description states: "Lowering below 5Hz(default) is not allowed." The EKF scales buffers based on this rate.
- **Source**: Source #1 (ArduPilot AP_GPS_Params)
- **Phase**: Assessment
- **Target Audience**: ArduPilot-based flight controllers
- **Confidence**: ✅ High
- **Related Dimension**: Flight controller integration completeness
-
-## Fact #2 — GPS_INPUT requires velocity + accuracy + GPS time fields
- **Statement**: GPS_INPUT message requires not just lat/lon/alt, but also: vn/ve/vd velocity components, hdop/vdop, horiz_accuracy/vert_accuracy/speed_accuracy, fix_type, time_week/time_week_ms, satellites_visible, and ignore_flags bitmap. The solution draft05 mentions GPS_INPUT but does not specify how these fields are populated (especially velocity from ESKF, accuracy from covariance, GPS time conversion from system time).
- **Source**: Source #2 (MAVLink GPS_INPUT definition), Source #3 (pymavlink example)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Flight controller integration completeness
-
-## Fact #3 — ESKF state vector needs explicit definition
- **Statement**: Standard ESKF for UAV VIO fusion uses 15-16 state error vector: δp[3], δv[3], δθ[3] (attitude error in so(3)), δba[3] (accel bias), δbg[3] (gyro bias), optionally δg[3] (gravity). The solution draft05 says "16-state vector" and "ESKF + buffers ~10MB" but never defines the actual state vector, process model (F, Q matrices), measurement models (H matrices for VO and satellite), or noise parameters.
- **Source**: Source #5 (ludvigls/ESKF), Source #6 (EliaTarasov/ESKF)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Sensor fusion completeness
-
-## Fact #4 — Monocular VIO has scale ambiguity without excitation
- **Statement**: Monocular visual-inertial odometry cannot observe metric scale during constant-velocity flight (zero accelerometer excitation). This is a fundamental observability limitation. The solution uses monocular cuVSLAM + IMU, and fixed-wing UAVs fly mostly at constant velocity. Scale must be provided externally — via known altitude (barometric + predefined mission altitude) or satellite matching absolute position.
- **Source**: Source #7 (Range-VIO scale observability paper)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Sensor fusion completeness, core algorithm correctness
-
-## Fact #5 — cuVSLAM requires explicit camera calibration and IMU calibration
- **Statement**: PyCuVSLAM requires Camera(fx, fy, cx, cy, width, height) + Distortion model + ImuCalibration(gyroscope_noise_density, gyroscope_random_walk, accelerometer_noise_density, accelerometer_random_walk, frequency, T_imu_rig). The solution draft05 does not specify any camera calibration procedure, IMU noise parameters, or the T_imu_rig (IMU-to-camera) extrinsic transformation.
- **Source**: Source #4 (PyCuVSLAM docs)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Visual odometry completeness
-
-## Fact #6 — cuVSLAM IMU fallback provides ~1s acceptable tracking
- **Statement**: When visual tracking fails (featureless terrain, darkness), cuVSLAM falls back to IMU-only integration which provides "approximately 1 second" of acceptable tracking quality before drift becomes unacceptable. After that, tracking is lost.
- **Source**: Source #4 (PyCuVSLAM/Isaac ROS docs)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Resilience, edge case handling
-
-## Fact #7 — Disconnected route segments need satellite re-localization
- **Statement**: When a UAV makes a sharp turn and the next photos have no overlap with previous frames, cuVSLAM will lose tracking. The solution must re-localize using satellite imagery. The AC requires handling "more than 2 such disconnected segments" as a core strategy. Solution draft05 mentions this requirement but does not define the concrete re-localization algorithm (how satellite match triggers, how the new position is initialized in ESKF, how the map is connected to the previous segment).
- **Source**: Source #8 (NaviLoc), Source #9 (SatLoc-Fusion), AC requirements
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Functional completeness — disconnected segments
-
-## Fact #8 — Coordinate transformation chain is undefined
- **Statement**: The system needs a well-defined coordinate transformation chain: (1) pixel coordinates → camera frame (using intrinsics), (2) camera frame → body frame (camera mount extrinsics), (3) body frame → NED frame (using attitude from ESKF), (4) NED → WGS84 (using reference point). For satellite matching: geo-referenced tile coordinates → WGS84. For object localization: pixel + camera angle + altitude → ground point → WGS84. None of these transformations are explicitly defined in draft05.
- **Source**: Source #15 (Istanbul Tech thesis), Source #7
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Coordinate system completeness
-
-## Fact #9 — GSD normalization required for satellite-aerial matching
- **Statement**: Camera GSD at 600m altitude with ADTI 20L V1 (16mm, APS-C) is ~15.9 cm/pixel. Google Maps zoom 19 ≈ 0.3 m/pixel, zoom 18 ≈ 0.6 m/pixel. The GSD ratio is ~2:1 to ~4:1 depending on zoom level and altitude. Draft05's "pre-resize" step in the offline pipeline is mentioned but not specified: what resolution? what zoom level? The matching model (LiteSAM/XFeat) input size must match appropriately.
- **Source**: Source #14 (GSD matching paper), solution_draft05 calculations
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Satellite matching completeness
-
-## Fact #10 — Google Maps imagery in conflict zones is intentionally outdated
- **Statement**: Google Maps deliberately serves older imagery (>1 year) for conflict zones in Ukraine. Ukrainian officials have flagged security concerns. The operational area (eastern/southern Ukraine) is directly in the conflict zone. Imagery may be 1-3+ years old, with seasonal differences (summer tiles vs winter flight, or vice versa). This is a HIGH-severity gap for satellite matching accuracy.
- **Source**: Source #12 (Google Maps Ukraine coverage)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Satellite imagery quality risk
-
-## Fact #11 — No operational procedures defined
- **Statement**: Mature GPS-denied systems (Auterion, PX4) define: pre-flight checklist (set home position, verify sensors, verify tile coverage), in-flight monitoring procedures (what to watch, when to intervene), and post-flight analysis (compare estimated vs actual GPS on return). Solution draft05 has no operational procedures section.
- **Source**: Source #10 (Auterion), Source #11 (PX4)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system operators
- **Confidence**: ✅ High
- **Related Dimension**: Operational maturity
-
-## Fact #12 — Object localization lacks implementation detail
- **Statement**: AC requires: "Other onboard AI systems can request GPS coordinates of objects detected by the AI camera." The solution says "trigonometric calculation using UAV GPS position, camera angle, zoom, altitude." But no API is defined, no coordinate math is shown, no handling of camera zoom/angle → ground projection is specified. The Viewpro A40 Pro gimbal angle and zoom parameters are not integrated.
- **Source**: Acceptance criteria, solution_draft05
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Object localization completeness
-
-## Fact #13 — Tech stack document is inconsistent with draft05
- **Statement**: tech_stack.md says "camera @ ~3fps" in non-functional requirements. Draft05 corrected this to 0.7fps. tech_stack.md lists LiteSAM benchmark decision at 480px/640px/800px; draft05 uses 1280px. tech_stack.md doesn't mention EfficientLoFTR as fallback. These inconsistencies indicate the tech_stack.md was not updated after draft05 changes.
- **Source**: tech_stack.md, solution_draft05.md
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Document consistency, maturity
-
-## Fact #14 — Confidence scoring is undefined in draft05
- **Statement**: Draft05 says "Confidence Scoring → GPS_INPUT Mapping — Unchanged from draft03" but draft05 is supposed to be self-contained. The actual confidence scoring logic (how VO confidence + satellite match confidence map to GPS_INPUT fix_type, hdop, horiz_accuracy) is never defined in the current draft. This is critical because ArduPilot's EKF uses these accuracy fields to weight the GPS data.
- **Source**: Source #2 (GPS_INPUT fields), solution_draft05
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Flight controller integration, confidence scoring
-
-## Fact #15 — Initial bootstrap sequence is incomplete
- **Statement**: Draft05 startup reads GLOBAL_POSITION_INT to get initial GPS position. But: (1) cuVSLAM needs its first frame + features to initialize — how is the first satellite match triggered? (2) ESKF needs initial state — position from GPS, but velocity? attitude? (3) How does the system know GPS is denied and should start sending GPS_INPUT? (4) Is there a handoff protocol from real GPS to GPS-denied system?
- **Source**: solution_draft05, Source #10 (Auterion procedures)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Startup and bootstrap completeness
-
-## Fact #16 — No recovery from companion computer reboot
- **Statement**: AC requires: "On companion computer reboot mid-flight, the system should attempt to re-initialize from the flight controller's current IMU-extrapolated position." Draft05 does not address this scenario. The system needs: read current FC position estimate, re-initialize ESKF, reload TRT engines (~1-3s), start cuVSLAM with no prior map, trigger immediate satellite re-localization.
- **Source**: Acceptance criteria, solution_draft05
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Resilience, failsafe completeness
-
-## Fact #17 — No position refinement mechanism
- **Statement**: AC states: "The system may refine previously calculated positions and send corrections to the flight controller as updated estimates." Draft05 does not define how this works. When a satellite match provides an absolute correction, do previously estimated positions get retroactively corrected? Is this communicated to the flight controller? How?
- **Source**: Acceptance criteria, solution_draft05
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ⚠️ Medium (AC is ambiguous about necessity)
- **Related Dimension**: Position refinement
-
-## Fact #18 — Tile storage requirements not calculated
- **Statement**: The solution mentions "preload tiles ±2km" and "GeoHash-indexed directory" but never calculates: how many tiles are needed for a mission area, what storage space is required, what zoom levels to use, or how to handle the trade-off between tile coverage area and storage limit. At zoom 19 (~0.3m/pixel), each 256×256 tile covers ~77m × 77m. Covering a 200km flight path with ±2km buffer would require ~130,000 tiles (~2.5GB JPEG).
- **Source**: solution_draft05, tech_stack.md
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ⚠️ Medium (estimates based on standard tile sizes)
- **Related Dimension**: Offline preparation completeness
-
-## Fact #19 — 3 consecutive failed frames → re-localization request undefined
- **Statement**: AC requires: "If system cannot determine position of 3 consecutive frames by any means, send re-localization request to ground station operator via telemetry link." Draft05 does not define: (1) the re-localization request message format, (2) what "any means" includes (VO failed + satellite match failed + IMU drift exceeded threshold?), (3) how the operator response is received and applied, (4) what the system does while waiting.
- **Source**: Acceptance criteria, solution_draft05
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: Ground station integration, resilience
-
-## Fact #20 — FastAPI endpoints mentioned but not defined
- **Statement**: Draft05 mentions FastAPI for "local IPC" but the REST API endpoints are only defined in security_analysis.md (POST /sessions, GET /sessions/{id}/stream, POST /sessions/{id}/anchor, DELETE /sessions/{id}). The solution draft itself doesn't specify the API contract, request/response schemas, or how other onboard systems interact.
- **Source**: solution_draft05, security_analysis.md
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ✅ High
- **Related Dimension**: API completeness
-
-## Fact #21 — NaviLoc achieves 19.5m error with trajectory-level optimization
- **Statement**: Recent research (NaviLoc, 2025) shows that trajectory-level optimization treating satellite matching as noisy measurement achieves 19.5m mean error at 50-150m altitude, 16× better than per-frame matching. The solution's approach of per-keyframe satellite matching + ESKF correction is simpler but potentially less accurate than trajectory-level optimization.
- **Source**: Source #8 (NaviLoc)
- **Phase**: Assessment
- **Target Audience**: GPS-denied system
- **Confidence**: ⚠️ Medium (NaviLoc operates at lower altitude with higher overlap)
- **Related Dimension**: Algorithm maturity, accuracy potential
@@ -1,63 +0,0 @@
-# Comparison Framework
-
-## Selected Framework Type
-Knowledge Organization + Problem Diagnosis — mapping completeness dimensions against solution state
-
-## Completeness Dimensions
-
-1. Core Pipeline Definition
-2. Sensor Fusion (ESKF) Specification
-3. Visual Odometry Configuration
-4. Satellite Image Matching Pipeline
-5. Coordinate System & Transformations
-6. Flight Controller Integration (GPS_INPUT)
-7. Disconnected Route Segment Handling
-8. Startup, Bootstrap & Failsafe
-9. Object Localization
-10. Offline Preparation Pipeline
-11. Ground Station Integration
-12. Operational Procedures
-13. API & Inter-system Communication
-14. Document Consistency
-15. Testing Coverage vs AC
-
-## Maturity Scoring
-
-| Score | Level | Description |
-|-------|-------|-------------|
-| 1 | Concept | Mentioned but not specified |
-| 2 | Defined | Architecture-level description, component selected |
-| 3 | Detailed | Implementation details, data flows, algorithms specified |
-| 4 | Validated | Benchmarked, tested, edge cases handled |
-| 5 | Production | Field-tested, operational procedures, monitoring |
-
-## Completeness Assessment Matrix
-
-| Dimension | Current State | Maturity | Key Gaps | Facts |
-|-----------|--------------|----------|----------|-------|
-| Core Pipeline | VO → ESKF → GPS_INPUT flow well-defined. Dual CUDA stream architecture. Camera rate corrected to 0.7fps. Time budgets calculated. | 3 | Self-contained — but references "unchanged from draft03" in several places instead of restating | — |
-| ESKF Specification | "16-state vector", "ESKF + buffers ~10MB", "ESKF measurement update" | 1.5 | No state vector definition, no process model (F,Q), no measurement models (H for VO, H for satellite), no noise parameters, no scale observability analysis, no tuning strategy | #3, #4 |
-| VO Configuration | cuVSLAM selected, 0.7fps feasibility analyzed, pyramid-LK range calculated, overlap >95% | 2.5 | No camera calibration procedure, no IMU calibration parameters (noise density, random walk), no T_imu_rig extrinsic, no cuVSLAM mode selection (Mono vs Inertial), no cuVSLAM initialization procedure | #5, #6 |
-| Satellite Matching | LiteSAM/EfficientLoFTR/XFeat decision tree, TRT conversion workflow, async Stream B | 2.5 | No GSD normalization spec, no tile-to-camera scale matching, no matching confidence threshold, no geometric verification details beyond "RANSAC homography", no match-to-WGS84 conversion | #9, #14 |
-| Coordinate System | WGS84 output mentioned. Camera footprint calculated. | 1 | No transformation chain defined (pixel→camera→body→NED→WGS84). No camera-to-body extrinsic. No reference point definition for NED. No handling of terrain elevation. | #8 |
-| Flight Controller Integration | pymavlink, GPS_INPUT, 5-10Hz, UART | 2 | No GPS_INPUT field population spec (where do velocity, accuracy, hdop come from?). No fix_type mapping. No GPS time conversion. No ignore_flags. Confidence scoring undefined in current draft. | #1, #2, #14 |
-| Disconnected Segments | Mentioned in AC, satellite matching acknowledged as solution | 1.5 | No algorithm for detecting tracking loss. No re-localization trigger. No position initialization after re-localization. No map discontinuity handling. | #7 |
-| Startup & Failsafe | 12-step startup sequence. Engine load times. | 2 | No GPS-denied handoff protocol. No mid-flight reboot recovery. No "3 consecutive failed frames" handling. No operator re-localization workflow. | #15, #16, #19 |
-| Object Localization | "Trigonometric calculation" mentioned | 1 | No math defined. No API endpoint. No Viewpro gimbal integration spec. No accuracy analysis. | #12 |
-| Offline Preparation | Tile download → validate → pre-resize → store. TRT engine build. | 2 | No zoom level selection. No storage calculation. No coverage verification. No tile freshness check. No pre-flight validation tool. | #18 |
-| Ground Station Integration | NAMED_VALUE_FLOAT at 1Hz for confidence/drift. Operator re-localization hint mentioned in AC. | 1.5 | Re-localization request/response undefined. Ground station display requirements undefined. Operator workflow undefined. | #19 |
-| Operational Procedures | None defined | 0 | No pre-flight checklist. No in-flight monitoring guide. No post-flight analysis. No failure response procedures. | #11 |
-| API & IPC | FastAPI mentioned for "local IPC" | 1.5 | Endpoints only in security_analysis.md, not in solution. No request/response schemas. No SSE event format. No object localization API. | #20 |
-| Document Consistency | 3 documents (draft05, tech_stack, security) | — | tech_stack.md has 3fps (should be 0.7fps). LiteSAM resolution mismatch. EfficientLoFTR missing from tech_stack. | #13 |
-| Testing vs AC | Tests cover TRT, cuVSLAM 0.7fps, shutter. | 2.5 | No explicit mapping of tests to AC items. Missing tests: disconnected segments, re-localization, 3-consecutive-failure, object localization, operator workflow, mid-flight reboot. | — |
-
-## Overall Maturity Assessment
-
-| Category | Avg Score | Assessment |
-|----------|-----------|------------|
-| Hardware/Platform | 3.5 | Well-researched: UAV specs, camera analysis, memory budget, thermal |
-| Core Algorithms (VO, matching) | 2.5 | Component selection solid, but implementation specs missing |
-| Sensor Fusion (ESKF) | 1.5 | Severely under-specified |
-| System Integration | 1.5 | GPS_INPUT, coordinate transforms, API all incomplete |
-| Operational Readiness | 0.5 | No operational procedures, no deployment pipeline |
-| **Overall** | **~2.0** | **Architecture-level design, not implementation-ready** |
@@ -1,166 +0,0 @@
-# Reasoning Chain
-
-## Dimension 1: ESKF Sensor Fusion Specification
-
-### Fact Confirmation
-Per Fact #3, standard ESKF for UAV VIO uses 15-16 state error vector: δp[3], δv[3], δθ[3], δba[3], δbg[3]. Per Fact #4, monocular VIO cannot observe metric scale during constant-velocity flight (fundamental observability limitation). Per Fact #7, scale requires external constraint (altitude or satellite absolute position).
-
-### Current State
-Draft05 says "Custom ESKF (NumPy/SciPy)", "16-state vector", "ESKF measurement update ~1ms", "ESKF IMU prediction at 5-10Hz". But provides zero mathematical detail.
-
-### Conclusion
-The ESKF is the **most under-specified critical component**. Without defining:
- State vector and error state vector explicitly
- Process model (how IMU data propagates the state)
- VO measurement model (how cuVSLAM relative pose updates the filter)
- Satellite measurement model (how absolute position corrections are applied)
- How scale is maintained (altitude constraint? satellite corrections only?)
- Q and R matrices (at least initial values and tuning approach)
-
-...the system cannot be implemented. The ESKF is the central hub connecting all sensors — its specification drives the entire data flow.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 2: Flight Controller Integration (GPS_INPUT)
-
-### Fact Confirmation
-Per Fact #1, ArduPilot requires minimum 5Hz GPS_INPUT rate. Per Fact #2, GPS_INPUT has 15+ mandatory fields including velocity, accuracy, fix_type, GPS time. Per Fact #14, confidence scoring that maps internal state to GPS_INPUT accuracy fields is undefined.
-
-### Current State
-Draft05 specifies: pymavlink, GPS_INPUT, 5-10Hz, UART. This satisfies the rate requirement. But the message population is unspecified.
-
-### Conclusion
-The GPS_INPUT integration has the right architecture (5-10Hz, pymavlink, UART) but is missing the **data mapping layer**:
- `vn, ve, vd` must come from ESKF velocity estimate — requires ESKF to output velocity
- `horiz_accuracy, vert_accuracy` must come from ESKF covariance matrix (sqrt of position covariance diagonal)
- `hdop, vdop` need to be synthesized from accuracy values (hdop ≈ horiz_accuracy / expected_CEP_factor)
- `fix_type` must map from internal confidence (3=3D fix when satellite-anchored, 2=2D when VO-only?)
- `speed_accuracy` from ESKF velocity covariance
- GPS time (time_week, time_week_ms) requires conversion from system time to GPS epoch
- `satellites_visible` should be set to a constant (e.g., 10) to avoid triggering satellite-count failsafes
-
-This is a tractable implementation detail but must be specified before coding.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 3: Coordinate System & Transformations
-
-### Fact Confirmation
-Per Fact #8, the system needs pixel → camera → body → NED → WGS84 chain. Per Fact #9, satellite tiles have different GSD than camera imagery. Per Source #15, similar systems define explicit coordinate transforms with PnP solvers.
-
-### Current State
-Draft05 calculates camera footprint and GSD but never defines the transformation chain. Object localization mentions "trigonometric calculation" without math.
-
-### Conclusion
-This is a **fundamental architectural gap**. Every position estimate flows through coordinate transforms. Without defining them:
- cuVSLAM outputs relative pose in camera frame — how is this converted to NED displacement?
- Satellite matching outputs pixel correspondences — how does homography → WGS84 position?
- Object localization needs camera ray → ground intersection — impossible without camera-to-body and body-to-NED transforms
- The camera is "not autostabilized" — so body frame attitude matters for ground projection
-
-The fix requires defining: camera intrinsic matrix K, camera-to-body rotation T_cam_body, and the ESKF attitude estimate for body-to-NED.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 4: Disconnected Route Segments
-
-### Fact Confirmation
-Per Fact #7, AC explicitly requires handling disconnected segments as "core to the system." Per Fact #6, cuVSLAM IMU fallback gives ~1s before tracking loss. Per Source #8, trajectory-level optimization can handle segment connections.
-
-### Current State
-Draft05 acknowledges this in AC but the solution section says "sharp-turn frames are expected to fail VO and should be handled by satellite-based re-localization." No algorithm is specified.
-
-### Conclusion
-The solution needs a concrete **re-localization pipeline**:
-1. Detect tracking loss (cuVSLAM returns tracking_lost state)
-2. Continue ESKF with IMU-only prediction (high uncertainty growth)
-3. Immediately trigger satellite matching on next available frame
-4. If satellite match succeeds: reset ESKF position to matched position, reset cuVSLAM (or start new track)
-5. If satellite match fails: retry on next frame, increment failure counter
-6. If 3 consecutive failures: send re-localization request to ground station
-7. When new segment starts: mark as disconnected, continue building trajectory
-8. Optionally: if a later satellite match connects two segments to the same reference, merge them
-
-This is not trivial but follows directly from the existing architecture. It's a missing algorithm, not a missing component.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 5: Startup Bootstrap & Failsafe
-
-### Fact Confirmation
-Per Fact #15, the bootstrap sequence has gaps (first satellite match, initial ESKF state, GPS-denied handoff). Per Fact #16, AC requires mid-flight reboot recovery. Per Fact #19, AC requires 3-consecutive-failure re-localization request.
-
-### Current State
-Draft05 has a 12-step startup sequence that covers the happy path. The failure paths and special cases are not addressed.
-
-### Conclusion
-Three failsafe scenarios need specification:
-1. **GPS-denied handoff**: How does the system know to start? Options: (a) always running — takes over when GPS quality degrades, (b) operator command, (c) automatic GPS quality monitoring. The system should probably always be running in parallel and the FC uses the best available source.
-2. **Mid-flight reboot**: Read FC position → init ESKF with high uncertainty → start cuVSLAM → immediate satellite match → within ~5s should have a position estimate. TRT engine load (1-3s) is the main startup cost.
-3. **3 consecutive failures**: Define "failure" precisely (VO lost + satellite match failed + IMU-only drift > threshold). Send NAMED_VALUE_FLOAT or custom MAVLink message to ground station. Define operator response format.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 6: Satellite Matching Pipeline Details
-
-### Fact Confirmation
-Per Fact #9, camera GSD is ~15.9 cm/pixel at 600m with ADTI+16mm lens. Satellite at zoom 19 is ~0.3 m/pixel. Per Fact #10, Google Maps imagery in Ukraine conflict zone is intentionally >1 year old. Per Fact #14, the solution says "pre-resize" but doesn't specify to what resolution.
-
-### Current State
-Draft05 has a solid model selection decision tree (LiteSAM → EfficientLoFTR → XFeat) and TRT conversion workflow. But the actual matching pipeline data flow is incomplete.
-
-### Conclusion
-The matching pipeline needs:
- **Input preparation**: Camera frame (5456×3632 at ~15.9 cm/pixel at 600m) → downsample to matcher input resolution (1280px for LiteSAM). Satellite tile at zoom 18 (~0.6 m/pixel) → no resize needed if using 256px tiles, or assemble 5×5 tile mosaic for coverage.
- **GSD matching**: Either downsample camera image to satellite GSD, or specify that the matcher handles multi-scale internally. LiteSAM was designed for satellite-aerial matching so it may handle this. XFeat is general-purpose and may need explicit scale normalization.
- **Tile selection**: Given ESKF position estimate + uncertainty, select the correct satellite tile(s). What if the position estimate has drifted and the wrong tile is selected? Need a search radius based on ESKF covariance.
- **Match → position**: Homography from RANSAC → decompose to get translation in satellite coordinate frame → convert to WGS84 using tile's geo-reference.
- **Seasonal/temporal mismatch**: Tiles could be from different seasons. Feature matching must be robust to appearance changes.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 7: Operational Maturity
-
-### Fact Confirmation
-Per Fact #11, mature systems (Auterion, PX4) define pre-flight checklists, in-flight monitoring, failure response procedures. Per Fact #13, documents are inconsistent (tech_stack.md still says 3fps).
-
-### Current State
-Zero operational procedures defined. Documents are partially inconsistent.
-
-### Conclusion
-This is expected at this stage of development (architecture/design phase). Operational procedures should come after implementation and initial testing. However, the document inconsistencies should be fixed now to avoid confusion during implementation. The tech_stack.md and solution_draft should be aligned.
-
-### Confidence: ✅ High
-
---
-
-## Dimension 8: Object Localization
-
-### Fact Confirmation
-Per Fact #12, AC requires other AI systems to request GPS coordinates of detected objects. The Viewpro A40 Pro gimbal has configurable angle and zoom.
-
-### Current State
-Draft05 says "trigonometric calculation using UAV GPS position, camera angle, zoom, altitude. Flat terrain assumed."
-
-### Conclusion
-The math is straightforward but needs specification:
- Input: pixel coordinates (u,v) in Viewpro image, gimbal angles (pan, tilt), zoom level, UAV position (from GPS-denied system), UAV altitude
- Process: (1) pixel → ray in camera frame using intrinsics + zoom, (2) camera frame → body frame using gimbal angles, (3) body frame → NED using UAV attitude, (4) ray-ground intersection assuming flat terrain at known altitude, (5) NED offset → WGS84
- Output: lat, lon of object + accuracy estimate (propagated from UAV position accuracy + gimbal angle uncertainty)
- API: FastAPI endpoint for other onboard systems to call
-
-This is a 2-point complexity task but should be specified in the solution.
-
-### Confidence: ✅ High
@@ -1,96 +0,0 @@
-# Validation Log
-
-## Validation Scenario 1: Normal Straight Flight (Happy Path)
-
-### Expected Based on Conclusions
-UAV flies straight at 70 km/h, 600m altitude. ADTI captures at 0.7fps. cuVSLAM processes each frame (~9ms), ESKF fuses VO + IMU. Every 5-10 frames, satellite matching provides absolute correction. GPS_INPUT sent at 5-10Hz.
-
-### Actual Validation Results
-The happy path is well-specified in draft05. Time budgets, memory budgets, overlap calculations all valid. The 5-10Hz ESKF IMU prediction fills gaps between 0.7fps camera frames. Satellite matching async on Stream B.
-
-**GAP**: Even on straight flight, the GPS_INPUT message field population is undefined. Where does velocity come from? What fix_type is sent? What accuracy values?
-
---
-
-## Validation Scenario 2: Sharp Turn with No Overlap
-
-### Expected Based on Conclusions
-UAV makes a 90° turn. Next frame has zero overlap with previous. cuVSLAM loses tracking. System falls back to IMU. ESKF uncertainty grows rapidly. Satellite matching on next frame provides re-localization.
-
-### Actual Validation Results
-**CRITICAL GAP**: No algorithm defined for this scenario. Questions:
-1. How does the system detect cuVSLAM tracking loss? (cuVSLAM API presumably returns a tracking state)
-2. During IMU-only phase, what is the ESKF prediction uncertainty growth rate? (~1-2m/s drift with consumer IMU)
-3. When satellite match succeeds after the turn, how is ESKF re-initialized? (measurement update with very high innovation? or state reset?)
-4. How is cuVSLAM re-initialized on the new heading? (new track from scratch? or cuVSLAM loop closure if it sees previous terrain?)
-5. If the turn area is over featureless terrain (farmland), satellite matching may also fail — then what?
-
-The AC says "sharp-turn frames should be within 200m drift and angle <70 degrees" — this bounds the problem but the solution doesn't address it algorithmically.
-
---
-
-## Validation Scenario 3: Long Flight Over Uniform Terrain
-
-### Expected Based on Conclusions
-UAV flies 50km straight over large agricultural fields. cuVSLAM may struggle with low-texture terrain. Satellite matching is the only absolute correction source.
-
-### Actual Validation Results
-This scenario tests the limits of the design:
- cuVSLAM at 600m altitude sees ~577m × 870m footprint. If it's a single wheat field, features may be sparse. cuVSLAM falls back to IMU (~1s acceptable, then tracking lost).
- Satellite matching must carry the entire position estimation burden.
- At 0.7fps with keyframe every 5-10 frames: satellite match every 7-14s.
- Between matches, IMU-only drift at 70 km/h: ~14m/s × time × drift_rate. With consumer IMU: ~1-5m drift per second.
- Over 14s between matches: ~14-70m potential drift. AC requires <100m between anchors.
-
-**GAP**: The system's behavior in this degraded mode needs explicit specification. Is this VO-failed + satellite-only + IMU acceptable? What's the accuracy?
-
---
-
-## Validation Scenario 4: First Frame After GPS Denial
-
-### Expected Based on Conclusions
-GPS was working. Now it's denied/spoofed. System takes over.
-
-### Actual Validation Results
-**GAP**: No handoff protocol defined. Questions:
-1. How does the system detect GPS denial? (it doesn't — it's assumed the operator knows)
-2. The initial ESKF state comes from GLOBAL_POSITION_INT — but this might be the spoofed GPS. How to validate?
-3. If GPS is being spoofed rather than denied, the initial position could be wrong by kilometers.
-4. Draft05 assumes clean initial GPS. This is reasonable for the first version but should be acknowledged as a limitation.
-
---
-
-## Validation Scenario 5: Mid-Flight Companion Computer Reboot
-
-### Expected Based on Conclusions
-Companion computer crashes and reboots. Flight controller continues flying on IMU dead reckoning.
-
-### Actual Validation Results
-**GAP**: AC requires recovery. Draft05 doesn't address it. Sequence should be:
-1. Jetson boots (~30-60s depending on boot time)
-2. GPS-Denied service starts (systemd)
-3. Connect to FC, get current position (IMU-extrapolated, may have significant drift)
-4. Load TRT engines (~1-3s each, total ~2-6s)
-5. Start cuVSLAM (no prior map, fresh start)
-6. Immediate satellite match to get absolute position
-7. Total recovery time: ~35-70s. During this time, FC uses IMU-only. At 70 km/h: ~700-1400m of uncontrolled drift.
-
-This is a real operational concern and should be documented even if the solution is "acknowledge the limitation."
-
---
-
-## Counterexamples
- **NaviLoc achieves 19.5m at lower altitude**: Our system operates at 600-1000m with larger footprints and coarser GSD. The accuracy requirement (50m for 80%, 20m for 60%) is less demanding. The simpler ESKF approach may be adequate.
- **SatLoc-Fusion uses DinoV2 for place recognition**: Our system uses feature matching (LiteSAM/XFeat) which is more precise but less robust to appearance changes. DinoV2 is more robust to seasonal changes but gives coarser position.
-
-## Review Checklist
- [x] Draft conclusions consistent with fact cards
- [x] No important dimensions missed
- [x] No over-extrapolation
- [ ] Issue: ESKF specification is too underspecified to validate accuracy claims
- [ ] Issue: Disconnected segment handling is critical AC and has no algorithm
- [ ] Issue: GPS_INPUT field mapping undocumented
- [ ] Issue: Object localization API undefined
-
-## Conclusions Requiring Revision
-None requiring reversal. All identified gaps are genuine and supported by facts.
@@ -1,57 +0,0 @@
-# Question Decomposition
-
-## Original Question
-Should we switch from ONNX Runtime to native TensorRT Engine for all AI models in the GPS-Denied pipeline, running on Jetson Orin Nano Super?
-
-## Active Mode
-Mode B: Solution Assessment — existing solution_draft03.md uses ONNX Runtime / mixed inference. User requests focused investigation of TRT Engine migration.
-
-## Question Type
-Decision Support — evaluating a technology switch with cost/risk/benefit dimensions.
-
-## Research Subject Boundary
-
-| Dimension | Boundary |
-|-----------|----------|
-| Population | AI inference models in the GPS-Denied navigation pipeline |
-| Hardware | Jetson Orin Nano Super (8GB LPDDR5, 67 TOPS sparse INT8, 1020 MHz GPU, NO DLA) |
-| Software | JetPack 6.2 (CUDA 12.6, TensorRT 10.3, cuDNN 9.3) |
-| Timeframe | Current (2025-2026), JetPack 6.2 era |
-
-## AI Models in Pipeline
-
-| Model | Type | Current Runtime | TRT Applicable? |
-|-------|------|----------------|-----------------|
-| cuVSLAM | Native CUDA library (closed-source) | CUDA native | NO — already CUDA-optimized binary |
-| LiteSAM | PyTorch (MobileOne + TAIFormer + MinGRU) | Planned TRT FP16 | YES |
-| XFeat | PyTorch (learned features) | XFeatTensorRT exists | YES |
-| ESKF | Mathematical filter (Python/C++) | CPU/NumPy | NO — not an AI model |
-
-Only LiteSAM and XFeat are convertible to TRT Engine. cuVSLAM is already NVIDIA-native CUDA.
-
-## Decomposed Sub-Questions
-
-1. What is the performance difference between ONNX Runtime and native TRT Engine on Jetson Orin Nano Super?
-2. What is the memory overhead of ONNX Runtime vs native TRT on 8GB shared memory?
-3. What conversion paths exist for PyTorch → TRT Engine on Jetson aarch64?
-4. Are TRT engines hardware-specific? What's the deployment workflow?
-5. What are the specific conversion steps for LiteSAM and XFeat?
-6. Does Jetson Orin Nano Super have DLA for offloading?
-7. What are the risks and limitations of going TRT-only?
-
-## Timeliness Sensitivity Assessment
-
- **Research Topic**: TensorRT vs ONNX Runtime inference on Jetson Orin Nano Super
- **Sensitivity Level**: 🟠 High
- **Rationale**: TensorRT, JetPack, and ONNX Runtime release new versions frequently. Jetson Orin Nano Super mode is relatively new (JetPack 6.2, Jan 2025).
- **Source Time Window**: 12 months
- **Priority official sources to consult**:
-  1. NVIDIA TensorRT documentation (docs.nvidia.com)
-  2. NVIDIA JetPack 6.2 release notes
-  3. ONNX Runtime GitHub issues (microsoft/onnxruntime)
-  4. NVIDIA TensorRT GitHub issues (NVIDIA/TensorRT)
- **Key version information to verify**:
-  - TensorRT: 10.3 (JetPack 6.2)
-  - ONNX Runtime: 1.20.1+ (Jetson builds)
-  - JetPack: 6.2
-  - CUDA: 12.6
@@ -1,231 +0,0 @@
-# Source Registry
-
-## Source #1
- **Title**: ONNX Runtime Issue #24085: CUDA EP on Jetson Orin Nano does not use tensor cores
- **Link**: https://github.com/microsoft/onnxruntime/issues/24085
- **Tier**: L1 (Official GitHub issue with MSFT response)
- **Publication Date**: 2025-03-18
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: ONNX Runtime v1.20.1+, JetPack 6.1, CUDA 12.6
- **Target Audience**: Jetson Orin Nano developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: ONNX Runtime CUDA EP on Jetson Orin Nano is 7-8x slower than TRT standalone due to tensor cores not being utilized. Workaround: remove cudnn_conv_algo_search option and use FP16 models.
- **Related Sub-question**: Q1 (performance difference)
-
-## Source #2
- **Title**: ONNX Runtime Issue #20457: VRAM usage difference between TRT-EP and native TRT
- **Link**: https://github.com/microsoft/onnxruntime/issues/20457
- **Tier**: L1 (Official GitHub issue with MSFT dev response)
- **Publication Date**: 2024-04-25
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: ONNX Runtime 1.17.1, CUDA 12.2
- **Target Audience**: All ONNX Runtime + TRT users
- **Research Boundary Match**: ✅ Full match
- **Summary**: ONNX Runtime TRT-EP keeps serialized engine in memory (~420-440MB) during execution. Native TRT drops to 130-140MB after engine build by calling releaseBlob(). Delta: ~280-300MB.
- **Related Sub-question**: Q2 (memory overhead)
-
-## Source #3
- **Title**: ONNX Runtime Issue #12083: TensorRT Provider vs TensorRT Native
- **Link**: https://github.com/microsoft/onnxruntime/issues/12083
- **Tier**: L2 (Official MSFT dev response)
- **Publication Date**: 2022-07-05 (confirmed still relevant)
- **Timeliness Status**: ⚠️ Needs verification (old but fundamental architecture hasn't changed)
- **Version Info**: General ONNX Runtime
- **Target Audience**: All ONNX Runtime users
- **Research Boundary Match**: ✅ Full match
- **Summary**: MSFT engineer confirms TRT-EP "can achieve performance parity with native TensorRT." Benefit is automatic fallback for unsupported ops.
- **Related Sub-question**: Q1 (performance difference)
-
-## Source #4
- **Title**: ONNX Runtime Issue #11356: Lower performance on InceptionV3/4 with TRT EP
- **Link**: https://github.com/microsoft/onnxruntime/issues/11356
- **Tier**: L4 (Community report)
- **Publication Date**: 2022
- **Timeliness Status**: ⚠️ Needs verification
- **Version Info**: ONNX Runtime older version
- **Target Audience**: ONNX Runtime users
- **Research Boundary Match**: ⚠️ Partial (different model, but same mechanism)
- **Summary**: Reports ~3x performance difference (41 vs 129 inferences/sec) between ONNX RT TRT-EP and native TRT on InceptionV3/4.
- **Related Sub-question**: Q1 (performance difference)
-
-## Source #5
- **Title**: NVIDIA JetPack 6.2 Release Notes
- **Link**: https://docs.nvidia.com/jetson/archives/jetpack-archived/jetpack-62/release-notes/index.html
- **Tier**: L1 (Official NVIDIA documentation)
- **Publication Date**: 2025-01
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: JetPack 6.2, TensorRT 10.3, CUDA 12.6, cuDNN 9.3
- **Target Audience**: Jetson developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: JetPack 6.2 includes TensorRT 10.3, enables Super Mode for Orin Nano (67 TOPS, 1020 MHz GPU, 25W).
- **Related Sub-question**: Q3 (conversion paths)
-
-## Source #6
- **Title**: NVIDIA Jetson Orin Nano Super Developer Kit Blog
- **Link**: https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/
- **Tier**: L2 (Official NVIDIA blog)
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: Orin Nano Super, 67 TOPS sparse INT8
- **Target Audience**: Jetson developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Super mode: GPU 1020 MHz (vs 635), 67 TOPS sparse (vs 40), memory bandwidth 102 GB/s (vs 68), power 25W. No DLA cores on Orin Nano.
- **Related Sub-question**: Q6 (DLA availability)
-
-## Source #7
- **Title**: Jetson Orin module comparison (Connect Tech)
- **Link**: https://connecttech.com/jetson/jetson-module-comparison
- **Tier**: L3 (Authoritative hardware vendor)
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Jetson hardware buyers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Confirms Orin Nano has NO DLA cores. Orin NX has 1-2 DLA. AGX Orin has 2 DLA.
- **Related Sub-question**: Q6 (DLA availability)
-
-## Source #8
- **Title**: TensorRT Engine hardware specificity (NVIDIA/TensorRT Issue #1920)
- **Link**: https://github.com/NVIDIA/TensorRT/issues/1920
- **Tier**: L1 (Official NVIDIA TensorRT repo)
- **Publication Date**: 2022 (confirmed still valid for TRT 10)
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: All TensorRT versions
- **Target Audience**: TensorRT developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: TRT engines are tied to specific GPU models. Must build on target hardware. Cannot cross-compile x86→aarch64.
- **Related Sub-question**: Q4 (deployment workflow)
-
-## Source #9
- **Title**: trtexec ONNX to TRT conversion on Jetson Orin Nano (StackOverflow)
- **Link**: https://stackoverflow.com/questions/78787534/converting-a-pytorch-onnx-model-to-tensorrt-engine-jetson-orin-nano
- **Tier**: L4 (Community)
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Jetson developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Standard workflow: trtexec --onnx=model.onnx --saveEngine=model.trt --fp16. Use --memPoolSize instead of deprecated --workspace.
- **Related Sub-question**: Q3, Q5 (conversion workflow)
-
-## Source #10
- **Title**: TensorRT 10 Python API Documentation
- **Link**: https://docs.nvidia.com/deeplearning/tensorrt/10.15.1/inference-library/python-api-docs.html
- **Tier**: L1 (Official NVIDIA docs)
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: TensorRT 10.x
- **Target Audience**: TensorRT Python developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: TRT 10 uses tensor-based API (not binding indices). load engine via runtime.deserialize_cuda_engine(). Async inference via context.enqueue_v3(stream_handle).
- **Related Sub-question**: Q3 (conversion paths)
-
-## Source #11
- **Title**: Torch-TensorRT JetPack documentation
- **Link**: https://docs.pytorch.org/TensorRT/v2.10.0/getting_started/jetpack.html
- **Tier**: L1 (Official documentation)
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: Torch-TensorRT, JetPack 6.2, PyTorch 2.8.0
- **Target Audience**: PyTorch developers on Jetson
- **Research Boundary Match**: ✅ Full match
- **Summary**: Torch-TensorRT supports Jetson aarch64 with JetPack 6.2. Supports AOT compilation, FP16/INT8, dynamic shapes.
- **Related Sub-question**: Q3 (conversion paths)
-
-## Source #12
- **Title**: XFeatTensorRT GitHub repo
- **Link**: https://github.com/PranavNedunghat/XFeatTensorRT
- **Tier**: L4 (Community)
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: XFeat users on NVIDIA GPUs
- **Research Boundary Match**: ✅ Full match
- **Summary**: C++ TRT implementation of XFeat feature extraction. Already converts XFeat to TRT engine.
- **Related Sub-question**: Q5 (XFeat conversion)
-
-## Source #13
- **Title**: TensorRT Best Practices (Official NVIDIA)
- **Link**: https://docs.nvidia.com/deeplearning/tensorrt/latest/performance/best-practices.html
- **Tier**: L1 (Official NVIDIA docs)
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Version Info**: TensorRT 10.x
- **Target Audience**: TensorRT developers
- **Research Boundary Match**: ✅ Full match
- **Summary**: Comprehensive guide: use trtexec for benchmarking, --fp16 for FP16, use ModelOptimizer for INT8, use polygraphy for model inspection.
- **Related Sub-question**: Q3 (conversion workflow)
-
-## Source #14
- **Title**: NVIDIA blog: Maximizing DL Performance on Jetson Orin with DLA
- **Link**: https://developer.nvidia.com/blog/maximizing-deep-learning-performance-on-nvidia-jetson-orin-with-dla/
- **Tier**: L2 (Official NVIDIA blog)
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Jetson Orin developers (NX and AGX)
- **Research Boundary Match**: ⚠️ Partial (DLA not available on Orin Nano)
- **Summary**: DLA contributes 38-74% of total DL performance on Orin (NX/AGX). Supports CNN layers in FP16/INT8. NOT available on Orin Nano.
- **Related Sub-question**: Q6 (DLA availability)
-
-## Source #15
- **Title**: PUT Vision Lab: TensorRT vs ONNXRuntime comparison on Jetson
- **Link**: https://putvision.github.io/article/2021/12/20/jetson-onnxruntime-tensorrt.html
- **Tier**: L3 (Academic lab blog)
- **Publication Date**: 2021 (foundational comparison, architecture unchanged)
- **Timeliness Status**: ⚠️ Needs verification (older, but core findings still apply)
- **Target Audience**: Jetson developers
- **Research Boundary Match**: ⚠️ Partial (older Jetson, but same TRT vs ONNX RT question)
- **Summary**: Native TRT generally faster. ONNX RT TRT-EP adds wrapper overhead. Both use same TRT kernels internally.
- **Related Sub-question**: Q1 (performance difference)
-
-## Source #16
- **Title**: LiteSAM paper — MinGRU details (Eqs 12-16, Section 3.4.2)
- **Link**: https://www.mdpi.com/2072-4292/17/19/3349
- **Tier**: L1 (Peer-reviewed paper)
- **Publication Date**: 2025
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Satellite-aerial matching researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: MinGRU subpixel refinement uses 4 stacked layers, 3×3 window (9 candidates). Gates depend only on input C_f. Ops: Linear, Sigmoid, Mul, Add, ReLU, Tanh.
- **Related Sub-question**: Q5 (LiteSAM TRT compatibility)
-
-## Source #17
- **Title**: Coarse_LoFTR_TRT paper — LoFTR TRT adaptation for embedded devices
- **Link**: https://ar5iv.labs.arxiv.org/html/2202.00770
- **Tier**: L2 (arXiv paper with working open-source code)
- **Publication Date**: 2022
- **Timeliness Status**: ✅ Currently valid (TRT adaptation techniques still apply)
- **Target Audience**: Feature matching on embedded devices
- **Research Boundary Match**: ✅ Full match
- **Summary**: Documents specific code changes for TRT compatibility: einsum→elementary ops, ONNX export, knowledge distillation. Tested on Jetson Nano 2GB. 2.26M params reduced from 27.95M.
- **Related Sub-question**: Q5 (EfficientLoFTR as TRT-proven alternative)
-
-## Source #18
- **Title**: minGRU paper — "Were RNNs All We Needed?"
- **Link**: https://huggingface.co/papers/2410.01201
- **Tier**: L1 (Research paper)
- **Publication Date**: 2024-10
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: RNN/sequence model researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: MinGRU removes gate dependency on h_{t-1}, enabling parallel computation. Parallel implementation uses logcumsumexp for numerical stability. 175x faster than sequential for seq_len=512.
- **Related Sub-question**: Q5 (MinGRU TRT compatibility)
-
-## Source #19
- **Title**: SAM2 TRT performance degradation issue
- **Link**: https://github.com/facebookresearch/sam2/issues/639
- **Tier**: L4 (GitHub issue)
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: SAM/transformer TRT deployers
- **Research Boundary Match**: ⚠️ Partial (different model, but relevant for transformer attention TRT risks)
- **Summary**: SAM2 MemoryAttention 30ms PyTorch → 100ms TRT FP16. RoPEAttention bottleneck. Warning for transformer TRT conversion.
- **Related Sub-question**: Q7 (TRT conversion risks)
-
-## Source #20
- **Title**: EfficientLoFTR (CVPR 2024)
- **Link**: https://github.com/zju3dv/EfficientLoFTR
- **Tier**: L1 (CVPR paper + open-source code)
- **Publication Date**: 2024
- **Timeliness Status**: ✅ Currently valid
- **Target Audience**: Feature matching researchers
- **Research Boundary Match**: ✅ Full match
- **Summary**: 2.5x faster than LoFTR, higher accuracy. 15.05M params. Semi-dense matching. Available on HuggingFace under Apache 2.0. 964 GitHub stars.
- **Related Sub-question**: Q5 (alternative satellite matcher)
@@ -1,193 +0,0 @@
-# Fact Cards
-
-## Fact #1
- **Statement**: ONNX Runtime CUDA Execution Provider on Jetson Orin Nano (JetPack 6.1) is 7-8x slower than TensorRT standalone due to tensor cores not being utilized with default settings.
- **Source**: Source #1 (https://github.com/microsoft/onnxruntime/issues/24085)
- **Phase**: Assessment
- **Target Audience**: Jetson Orin Nano developers using ONNX Runtime
- **Confidence**: ✅ High (confirmed by issue author with NSight profiling, MSFT acknowledged)
- **Related Dimension**: Performance
-
-## Fact #2
- **Statement**: The workaround for Fact #1 is to remove the `cudnn_conv_algo_search` option (which defaults to EXHAUSTIVE) and use FP16 models. This restores tensor core usage.
- **Source**: Source #1
- **Phase**: Assessment
- **Target Audience**: Jetson Orin Nano developers
- **Confidence**: ✅ High (confirmed fix by issue author)
- **Related Dimension**: Performance
-
-## Fact #3
- **Statement**: ONNX Runtime TRT-EP keeps serialized TRT engine in memory (~420-440MB) throughout execution. Native TRT via trtexec drops to 130-140MB after engine deserialization by calling releaseBlob().
- **Source**: Source #2 (https://github.com/microsoft/onnxruntime/issues/20457)
- **Phase**: Assessment
- **Target Audience**: All ONNX RT TRT-EP users, especially memory-constrained devices
- **Confidence**: ✅ High (confirmed by MSFT developer @chilo-ms with detailed explanation)
- **Related Dimension**: Memory consumption
-
-## Fact #4
- **Statement**: The ~280-300MB extra memory from ONNX RT TRT-EP (Fact #3) is because the serialized engine is retained across compute function calls for dynamic shape models. Native TRT releases it after deserialization.
- **Source**: Source #2
- **Phase**: Assessment
- **Target Audience**: Memory-constrained Jetson deployments
- **Confidence**: ✅ High (MSFT developer explanation)
- **Related Dimension**: Memory consumption
-
-## Fact #5
- **Statement**: MSFT engineer states "TensorRT EP can achieve performance parity with native TensorRT" — both use the same TRT kernels internally. Benefit of TRT-EP is automatic fallback for unsupported ops.
- **Source**: Source #3 (https://github.com/microsoft/onnxruntime/issues/12083)
- **Phase**: Assessment
- **Target Audience**: General
- **Confidence**: ⚠️ Medium (official statement but contradicted by real benchmarks in some cases)
- **Related Dimension**: Performance
-
-## Fact #6
- **Statement**: Real benchmark of InceptionV3/4 showed ONNX RT TRT-EP achieving ~41 inferences/sec vs native TRT at ~129 inferences/sec — approximately 3x performance gap.
- **Source**: Source #4 (https://github.com/microsoft/onnxruntime/issues/11356)
- **Phase**: Assessment
- **Target Audience**: CNN model deployers
- **Confidence**: ⚠️ Medium (community report, older ONNX RT version, model-specific)
- **Related Dimension**: Performance
-
-## Fact #7
- **Statement**: Jetson Orin Nano Super specs: 67 TOPS sparse INT8 / 33 TOPS dense, GPU at 1020 MHz, 8GB LPDDR5 shared, 102 GB/s bandwidth, 25W TDP. NO DLA cores.
- **Source**: Source #6, Source #7
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ✅ High (official NVIDIA specs)
- **Related Dimension**: Hardware constraints
-
-## Fact #8
- **Statement**: Jetson Orin Nano has ZERO DLA (Deep Learning Accelerator) cores. DLA is only available on Orin NX (1-2 cores) and AGX Orin (2 cores).
- **Source**: Source #7, Source #14
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ✅ High (official hardware specifications)
- **Related Dimension**: Hardware constraints
-
-## Fact #9
- **Statement**: TensorRT engines are tied to specific GPU models, not just architectures. Must be built on the target device. Cannot cross-compile from x86 to aarch64.
- **Source**: Source #8 (https://github.com/NVIDIA/TensorRT/issues/1920)
- **Phase**: Assessment
- **Target Audience**: TRT deployers
- **Confidence**: ✅ High (NVIDIA confirmed)
- **Related Dimension**: Deployment workflow
-
-## Fact #10
- **Statement**: Standard conversion workflow: PyTorch → ONNX (torch.onnx.export) → trtexec --onnx=model.onnx --saveEngine=model.engine --fp16. Use --memPoolSize instead of deprecated --workspace flag.
- **Source**: Source #9, Source #13
- **Phase**: Assessment
- **Target Audience**: Model deployers on Jetson
- **Confidence**: ✅ High (official NVIDIA workflow)
- **Related Dimension**: Deployment workflow
-
-## Fact #11
- **Statement**: TensorRT 10.x Python API: load engine via runtime.deserialize_cuda_engine(data). Async inference via context.enqueue_v3(stream_handle). Uses tensor-name-based API (not binding indices).
- **Source**: Source #10
- **Phase**: Assessment
- **Target Audience**: Python TRT developers
- **Confidence**: ✅ High (official NVIDIA docs)
- **Related Dimension**: API/integration
-
-## Fact #12
- **Statement**: Torch-TensorRT supports Jetson aarch64 with JetPack 6.2. Supports ahead-of-time (AOT) compilation, FP16/INT8, dynamic and static shapes. Alternative path to ONNX→trtexec.
- **Source**: Source #11
- **Phase**: Assessment
- **Target Audience**: PyTorch developers on Jetson
- **Confidence**: ✅ High (official documentation)
- **Related Dimension**: Deployment workflow
-
-## Fact #13
- **Statement**: XFeatTensorRT repo exists — C++ TensorRT implementation of XFeat feature extraction. Confirms XFeat is TRT-convertible.
- **Source**: Source #12
- **Phase**: Assessment
- **Target Audience**: Our project (XFeat users)
- **Confidence**: ✅ High (working open-source implementation)
- **Related Dimension**: Model-specific conversion
-
-## Fact #14
- **Statement**: cuVSLAM is a closed-source NVIDIA CUDA library (PyCuVSLAM). It is NOT an ONNX or PyTorch model. It cannot and does not need to be converted to TRT — it's already native CUDA-optimized for Jetson.
- **Source**: cuVSLAM documentation (https://github.com/NVlabs/PyCuVSLAM)
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ✅ High (verified from PyCuVSLAM docs)
- **Related Dimension**: Model applicability
-
-## Fact #15
- **Statement**: JetPack 6.2 ships with TensorRT 10.3, CUDA 12.6, cuDNN 9.3. The tensorrt Python module is pre-installed and accessible on Jetson.
- **Source**: Source #5
- **Phase**: Assessment
- **Target Audience**: Jetson developers
- **Confidence**: ✅ High (official release notes)
- **Related Dimension**: Software stack
-
-## Fact #16
- **Statement**: TRT engine build on Jetson Orin Nano Super (8GB) can cause OOM for large models during the build phase, even if inference fits in memory. Workaround: build on a more powerful machine with same GPU architecture, or use Torch-TensorRT PyTorch workflow.
- **Source**: Source #5 (https://github.com/NVIDIA/TensorRT-LLM/issues/3149)
- **Phase**: Assessment
- **Target Audience**: Jetson Orin Nano developers building large TRT engines
- **Confidence**: ✅ High (confirmed in NVIDIA TRT-LLM issue)
- **Related Dimension**: Deployment workflow
-
-## Fact #17
- **Statement**: LiteSAM uses MobileOne backbone which is reparameterizable — multi-branch training structure collapses to a single feed-forward path. This is critical for TRT optimization: fewer layers, better fusion, faster inference.
- **Source**: Solution draft03, LiteSAM paper
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ✅ High (published paper)
- **Related Dimension**: Model-specific conversion
-
-## Fact #18
- **Statement**: INT8 quantization is safe for CNN layers (MobileOne backbone) but NOT for transformer components (TAIFormer in LiteSAM). FP16 is safe for both CNN and transformer layers.
- **Source**: Solution draft02/03 analysis
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ⚠️ Medium (general best practice, not verified on LiteSAM specifically)
- **Related Dimension**: Quantization strategy
-
-## Fact #19
- **Statement**: On 8GB shared memory Jetson: OS+runtime ~1.5GB, cuVSLAM ~200-500MB, tiles ~200MB. Remaining budget: ~5.8-6.1GB. ONNX RT TRT-EP overhead of ~280-300MB per model is significant. Native TRT saves this memory.
- **Source**: Solution draft03 memory budget + Source #2
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ✅ High (computed from verified facts)
- **Related Dimension**: Memory consumption
-
-## Fact #20
- **Statement**: LiteSAM's MinGRU subpixel refinement (Eqs 12-16) uses: z_t = σ(Linear(C_f)), h̃_t = Linear(C_f), h_t = (1-z_t)⊙h_{t-1} + z_t⊙h̃_t. Gates depend ONLY on input C_f (not h_{t-1}). Operates on 3×3 window (9 candidates), 4 stacked layers. All ops are standard: Linear, Sigmoid, Mul, Add, ReLU, Tanh.
- **Source**: LiteSAM paper (MDPI Remote Sensing, 2025, Eqs 12-16, Section 3.4.2)
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ✅ High (from the published paper)
- **Related Dimension**: LiteSAM TRT compatibility
-
-## Fact #21
- **Statement**: MinGRU's parallel implementation can use logcumsumexp (log-space parallel scan), which is NOT a standard ONNX operator. However, for seq_len=9 (LiteSAM's 3×3 window), a simple unrolled loop is equivalent and uses only standard ops.
- **Source**: minGRU paper + lucidrains/minGRU-pytorch implementation
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ⚠️ Medium (logcumsumexp risk depends on implementation; seq_len=9 makes rewrite trivial)
- **Related Dimension**: LiteSAM TRT compatibility
-
-## Fact #22
- **Statement**: EfficientLoFTR has a proven TRT conversion path via Coarse_LoFTR_TRT (138 stars). The paper documents specific code changes needed: replace einsum with elementary ops (view, bmm, reshape, sum), adapt for TRT-compatible functions. Tested on Jetson Nano 2GB (~5 FPS with distilled model).
- **Source**: Coarse_LoFTR_TRT paper (arXiv:2202.00770) + GitHub repo
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ✅ High (published paper + working open-source implementation)
- **Related Dimension**: Fallback satellite matcher
-
-## Fact #23
- **Statement**: EfficientLoFTR has 15.05M params (2.4x more than LiteSAM's 6.31M). On AGX Orin with PyTorch: ~620ms (LiteSAM is 19.8% faster). Semi-dense matching. CVPR 2024. Available on HuggingFace under Apache 2.0.
- **Source**: LiteSAM paper comparison + EfficientLoFTR docs
- **Phase**: Assessment
- **Target Audience**: Our project
- **Confidence**: ✅ High (published benchmarks)
- **Related Dimension**: Fallback satellite matcher
-
-## Fact #24
- **Statement**: SAM2's MemoryAttention showed performance DEGRADATION with TRT: 30ms PyTorch → 100ms TRT FP16. RoPEAttention identified as bottleneck. This warns that transformer attention modules may not always benefit from TRT conversion.
- **Source**: https://github.com/facebookresearch/sam2/issues/639
- **Phase**: Assessment
- **Target Audience**: Transformer model deployers
- **Confidence**: ⚠️ Medium (different model, but relevant warning for attention layers)
- **Related Dimension**: TRT conversion risks
@@ -1,38 +0,0 @@
-# Comparison Framework
-
-## Selected Framework Type
-Decision Support
-
-## Selected Dimensions
-1. Inference latency
-2. Memory consumption
-3. Deployment workflow complexity
-4. Operator coverage / fallback
-5. API / integration effort
-6. Hardware utilization (tensor cores)
-7. Maintenance / ecosystem
-8. Cross-platform portability
-
-## Comparison: Native TRT Engine vs ONNX Runtime (TRT-EP and CUDA EP)
-
-| Dimension | Native TRT Engine | ONNX Runtime TRT-EP | ONNX Runtime CUDA EP | Factual Basis |
-|-----------|-------------------|---------------------|----------------------|---------------|
-| Inference latency | Optimal — uses TRT kernels directly, hardware-tuned | Near-parity with native TRT (same kernels), but up to 3x slower on some models due to wrapper overhead | 7-8x slower on Orin Nano with default settings (tensor core issue) | Fact #1, #5, #6 |
-| Memory consumption | ~130-140MB after engine load (releases serialized blob) | ~420-440MB during execution (keeps serialized engine) | Standard CUDA memory + framework overhead | Fact #3, #4 |
-| Memory delta per model | Baseline | +280-300MB vs native TRT | Higher than TRT-EP | Fact #3, #19 |
-| Deployment workflow | PyTorch → ONNX → trtexec → .engine (must build ON target device) | PyTorch → ONNX → pass to ONNX Runtime session (auto-builds TRT engine) | PyTorch → ONNX → pass to ONNX Runtime session | Fact #9, #10 |
-| Operator coverage | Only TRT-supported ops. Unsupported ops = build failure | Auto-fallback to CUDA/CPU for unsupported ops | All ONNX ops supported via CUDA/cuDNN | Fact #5 |
-| API complexity | Lower-level: manual buffer allocation, CUDA streams, tensor management | Higher-level: InferenceSession, automatic I/O | Highest-level: same ONNX Runtime API | Fact #11 |
-| Hardware utilization | Full: tensor cores, layer fusion, kernel auto-tuning, mixed precision | Full TRT kernels for supported ops, CUDA fallback for rest | Broken on Orin Nano with default settings (no tensor cores) | Fact #1, #2 |
-| Maintenance | Engine must be rebuilt per TRT version and per GPU model | ONNX model is portable, engine rebuilt automatically | ONNX model is portable | Fact #9 |
-| Cross-platform | NVIDIA-only, hardware-specific engine files | Multi-platform ONNX model, TRT-EP only on NVIDIA | Multi-platform (NVIDIA, AMD, Intel, CPU) | Fact #9 |
-| Relevance to our project | ✅ Best — we deploy only on Jetson Orin Nano Super | ❌ Cross-platform benefit wasted — we're NVIDIA-only | ❌ Performance issue on our target hardware | Fact #7, #8 |
-
-## Per-Model Applicability
-
-| Model | Can Convert to TRT? | Recommended Path | Notes |
-|-------|---------------------|------------------|-------|
-| cuVSLAM | NO | N/A — already CUDA native | Closed-source NVIDIA library, already optimized |
-| LiteSAM | YES | PyTorch → reparameterize MobileOne → ONNX → trtexec --fp16 | INT8 safe for MobileOne backbone only, NOT TAIFormer |
-| XFeat | YES | PyTorch → ONNX → trtexec --fp16 (or use XFeatTensorRT C++) | XFeatTensorRT repo already exists |
-| ESKF | N/A | N/A — mathematical filter, not a neural network | Python/C++ NumPy |
@@ -1,124 +0,0 @@
-# Reasoning Chain
-
-## Dimension 1: Inference Latency
-
-### Fact Confirmation
-ONNX Runtime CUDA EP on Jetson Orin Nano is 7-8x slower than TRT standalone with default settings (Fact #1). Even with the workaround (Fact #2), ONNX RT adds wrapper overhead. ONNX RT TRT-EP claims "performance parity" (Fact #5), but real benchmarks show up to 3x gaps on specific models (Fact #6).
-
-### Reference Comparison
-Native TRT uses kernel auto-tuning, layer fusion, and mixed-precision natively — no framework wrapper. Our models (LiteSAM, XFeat) are CNN+transformer architectures where TRT's fusion optimizations are most impactful. LiteSAM's reparameterized MobileOne backbone (Fact #17) is particularly well-suited for TRT fusion.
-
-### Conclusion
-Native TRT Engine provides the lowest possible inference latency on Jetson Orin Nano Super. ONNX Runtime adds measurable overhead, ranging from negligible to 3x depending on model architecture and configuration. For our latency-critical pipeline (400ms total budget, satellite matching target ≤200ms), every millisecond matters.
-
-### Confidence
-✅ High — supported by multiple sources, confirmed NVIDIA optimization pipeline.
-
---
-
-## Dimension 2: Memory Consumption
-
-### Fact Confirmation
-ONNX RT TRT-EP keeps ~420-440MB during execution vs native TRT at ~130-140MB (Fact #3). This is ~280-300MB extra PER MODEL. On our 8GB shared memory Jetson, OS+runtime takes ~1.5GB, cuVSLAM ~200-500MB, tiles ~200MB (Fact #19).
-
-### Reference Comparison
-If we run both LiteSAM and XFeat via ONNX RT TRT-EP: ~560-600MB extra memory overhead. Via native TRT: this overhead drops to near zero.
-
-With native TRT:
- LiteSAM engine: ~50-80MB
- XFeat engine: ~30-50MB
-With ONNX RT TRT-EP:
- LiteSAM: ~50-80MB + ~280MB overhead = ~330-360MB
- XFeat: ~30-50MB + ~280MB overhead = ~310-330MB
-
-### Conclusion
-Native TRT saves ~280-300MB per model vs ONNX RT TRT-EP. On our 8GB shared memory device, this is 3.5-3.75% of total memory PER MODEL. With two models, that's ~7% of total memory saved — meaningful when memory pressure from cuVSLAM map growth is a known risk.
-
-### Confidence
-✅ High — confirmed by MSFT developer with detailed explanation of mechanism.
-
---
-
-## Dimension 3: Deployment Workflow
-
-### Fact Confirmation
-Native TRT requires: PyTorch → ONNX → trtexec → .engine file. Engine must be built ON the target Jetson device (Fact #9). Engine is tied to specific GPU model and TRT version. TRT engine build on 8GB Jetson can OOM for large models (Fact #16).
-
-### Reference Comparison
-ONNX Runtime auto-builds TRT engine from ONNX at first run (or caches). Simpler developer experience but first-run latency spike. Torch-TensorRT (Fact #12) offers AOT compilation as middle ground.
-
-Our models are small (LiteSAM 6.31M params, XFeat even smaller). Engine build OOM is unlikely for our model sizes. Build once before flight, ship .engine files.
-
-### Conclusion
-Native TRT requires an explicit offline build step (trtexec on Jetson), but this is a one-time cost per model version. For our use case (pre-flight preparation already includes satellite tile download), adding a TRT engine build to the preparation workflow is trivial. The deployment complexity is acceptable.
-
-### Confidence
-✅ High — well-documented workflow, our model sizes are small enough.
-
---
-
-## Dimension 4: Operator Coverage / Fallback
-
-### Fact Confirmation
-Native TRT fails if a model contains unsupported operators. ONNX RT TRT-EP auto-falls back to CUDA/CPU for unsupported ops (Fact #5). This is TRT-EP's primary value proposition.
-
-### Reference Comparison
-LiteSAM (MobileOne + TAIFormer + MinGRU) and XFeat use standard operations: Conv2d, attention, GRU, ReLU, etc. These are all well-supported by TensorRT 10.3. MobileOne's reparameterized form is pure Conv2d+BN — trivially supported. TAIFormer attention uses standard softmax/matmul — supported in TRT 10. MinGRU is a simplified GRU — may need verification.
-
-Risk: If any op in LiteSAM is unsupported by TRT, the entire export fails. Mitigation: verify with polygraphy before deployment. If an op fails, refactor or use Torch-TensorRT which can handle mixed TRT/PyTorch execution.
-
-### Conclusion
-For our specific models, operator coverage risk is LOW. Standard CNN+transformer ops are well-supported in TRT 10.3. ONNX RT's fallback benefit is insurance we're unlikely to need. MinGRU in LiteSAM should be verified, but standard GRU ops are TRT-supported.
-
-### Confidence
-⚠️ Medium — high confidence for MobileOne+TAIFormer, medium for MinGRU (needs verification on TRT 10.3).
-
---
-
-## Dimension 5: API / Integration Effort
-
-### Fact Confirmation
-Native TRT Python API (Fact #11): manual buffer allocation with PyCUDA, CUDA stream management, tensor setup via engine.get_tensor_name(). ONNX Runtime: simple InferenceSession with .run().
-
-### Reference Comparison
-TRT Python API requires ~30-50 lines of boilerplate per model (engine load, buffer allocation, inference loop). ONNX Runtime requires ~5-10 lines. However, this is write-once code, encapsulated in a wrapper class.
-
-Our pipeline already uses CUDA streams for cuVSLAM pipelining (Stream A for VO, Stream B for satellite matching). Adding TRT inference to Stream B is natural — just pass stream_handle to context.enqueue_v3().
-
-### Conclusion
-Slightly more code with native TRT, but it's boilerplate that gets written once and wrapped. The CUDA stream integration actually BENEFITS from native TRT — direct stream control enables better pipelining with cuVSLAM.
-
-### Confidence
-✅ High — well-documented API, straightforward integration.
-
---
-
-## Dimension 6: Hardware Utilization
-
-### Fact Confirmation
-ONNX RT CUDA EP does NOT use tensor cores on Jetson Orin Nano by default (Fact #1). Native TRT uses tensor cores, layer fusion, kernel auto-tuning automatically. Jetson Orin Nano Super has 16 tensor cores at 1020 MHz (Fact #7). No DLA available (Fact #8).
-
-### Reference Comparison
-Since there's no DLA to offload to, GPU is our only accelerator. Maximizing GPU utilization is critical. Native TRT squeezes every ounce from the 16 tensor cores. ONNX RT has a known bug preventing this on our exact hardware.
-
-### Conclusion
-Native TRT is the only way to guarantee full hardware utilization on Jetson Orin Nano Super. ONNX RT's tensor core issue (even if workaround exists) introduces fragility. Since we have no DLA, wasting GPU tensor cores is unacceptable.
-
-### Confidence
-✅ High — hardware limitation is confirmed, no alternative accelerator.
-
---
-
-## Dimension 7: Cross-Platform Portability
-
-### Fact Confirmation
-ONNX Runtime runs on NVIDIA, AMD, Intel, CPU. TRT engines are NVIDIA-specific and even GPU-model-specific (Fact #9).
-
-### Reference Comparison
-Our system deploys ONLY on Jetson Orin Nano Super. The companion computer is fixed hardware. There is no requirement or plan to run on non-NVIDIA hardware. Cross-platform portability has zero value for this project.
-
-### Conclusion
-ONNX Runtime's primary value proposition (portability) is irrelevant for our deployment. We trade unused portability for maximum performance and minimum memory usage.
-
-### Confidence
-✅ High — deployment target is fixed hardware.
@@ -1,65 +0,0 @@
-# Validation Log
-
-## Validation Scenario
-Full GPS-Denied pipeline running on Jetson Orin Nano Super (8GB) during a 50km flight with ~1500 frames at 3fps. Two AI models active: LiteSAM for satellite matching (keyframes) and XFeat as fallback. cuVSLAM running continuously for VO.
-
-## Expected Based on Conclusions
-
-### If using Native TRT Engine:
- LiteSAM TRT FP16 engine loaded: ~50-80MB GPU memory after deserialization
- XFeat TRT FP16 engine loaded: ~30-50MB GPU memory after deserialization
- Total AI model memory: ~80-130MB
- Inference runs on CUDA Stream B, directly integrated with cuVSLAM Stream A pipelining
- Tensor cores fully utilized at 1020 MHz
- LiteSAM satellite matching at estimated ~165-330ms (TRT FP16 at 1280px)
- XFeat matching at estimated ~50-100ms (TRT FP16)
- Engine files pre-built during offline preparation, stored on Jetson storage alongside satellite tiles
-
-### If using ONNX Runtime TRT-EP:
- LiteSAM via TRT-EP: ~330-360MB during execution
- XFeat via TRT-EP: ~310-330MB during execution
- Total AI model memory: ~640-690MB
- First inference triggers engine build (latency spike at startup)
- CUDA stream management less direct
- Same inference speed (in theory, per MSFT claim)
-
-### Memory budget comparison (total 8GB):
- Native TRT: OS 1.5GB + cuVSLAM 0.5GB + tiles 0.2GB + models 0.13GB + misc 0.1GB = ~2.43GB (30% used)
- ONNX RT TRT-EP: OS 1.5GB + cuVSLAM 0.5GB + tiles 0.2GB + models 0.69GB + ONNX RT overhead 0.15GB + misc 0.1GB = ~3.14GB (39% used)
- Delta: ~710MB (9% of total memory)
-
-## Actual Validation Results
-The memory savings from native TRT are confirmed by the mechanism explanation from MSFT (Source #2). The 710MB delta is significant given cuVSLAM map growth risk (up to 1GB on long flights without aggressive pruning).
-
-The workflow integration is validated: engine files can be pre-built as part of the existing offline tile preparation pipeline. No additional hardware or tools needed — trtexec is included in JetPack 6.2.
-
-## Counterexamples
-
-### Counterexample 1: MinGRU operator may not be supported in TRT
-MinGRU is a simplified GRU variant used in LiteSAM's subpixel refinement. Standard GRU is supported in TRT 10.3, but MinGRU may use custom operations. If MinGRU fails TRT export, options:
-1. Replace MinGRU with standard GRU (small accuracy loss)
-2. Split model: CNN+TAIFormer in TRT, MinGRU refinement in PyTorch
-3. Use Torch-TensorRT which handles mixed execution
-
-**Assessment**: Low risk. MinGRU is a simplification of GRU, likely uses subset of GRU ops.
-
-### Counterexample 2: Engine rebuild needed per TRT version update
-JetPack updates may change TRT version, invalidating cached engines. Must rebuild all engines after JetPack update.
-
-**Assessment**: Acceptable. JetPack updates are infrequent on deployed UAVs. Engine rebuild takes minutes.
-
-### Counterexample 3: Dynamic input shapes
-If camera resolution changes between flights, engine with static shapes must be rebuilt. Can use dynamic shapes in trtexec (--minShapes, --optShapes, --maxShapes) but at slight performance cost.
-
-**Assessment**: Acceptable. Camera resolution is fixed per deployment. Build engine for that resolution.
-
-## Review Checklist
- [x] Draft conclusions consistent with fact cards
- [x] No important dimensions missed
- [x] No over-extrapolation
- [x] Conclusions actionable/verifiable
- [x] Memory calculations verified against known budget
- [x] Workflow integration validated against existing offline preparation
-
-## Conclusions Requiring Revision
-None — all conclusions hold under validation.