gps-denied-onboard/docs/03_tests/36_sharp_turn_zero_overlap_spec.md

Acceptance Test: AC-4 - Sharp Turn Recovery

Summary

Validate Acceptance Criterion 4: "System should correctly continue the work even during sharp turns, where the next photo doesn't overlap at all, or overlaps in less than 5%. The next photo should be in less than 200m drift and at an angle of less than 70°"

Linked Acceptance Criteria

AC-4: Handle sharp turns with <5% overlap

Preconditions

1. System operational with L2 (Global Place Recognition) enabled
2. AnyLoc model and Faiss index ready
3. Test datasets:
   • Dataset A: AD000042, AD000044, AD000045, AD000046 (skip AD000043)
   • Dataset B: AD000003, AD000009 (5-frame gap)
4. Ground truth available

Test Description

Test system's ability to recover from "kidnapped robot" scenarios where sequential tracking fails due to zero overlap. Validates L2 global place recognition functionality.

Test Steps

Step 1: Create Sharp Turn Flight (Dataset A)

• Action: Create flight with AD000042, AD000044, AD000045, AD000046
• Expected Result: Flight created, gap in sequence detected

Step 2: Process Through L1

• Action: L1 processes AD000042
• Expected Result: AD000042 processed successfully

Step 3: Attempt Sequential Tracking (L1 Failure Expected)

• Action: L1 attempts AD000042 → AD000044 (skip AD000043)
• Expected Result:
  • L1 fails (overlap < 5% or zero)
  • Low inlier count (< 10 matches)
  • System triggers L2 recovery

Step 4: L2 Global Relocalization

• Action: L2 (AnyLoc) queries AD000044 against satellite database
• Expected Result:
  • L2 retrieves correct satellite tile region
  • Coarse location found (within 200m of ground truth per AC-4)
  • Top-5 recall succeeds

Step 5: L3 Metric Refinement

• Action: L3 (LiteSAM) refines location using satellite tile
• Expected Result:
  • Precise GPS estimate (< 50m error)
  • High confidence score

Step 6: Continue Processing

• Action: Process AD000045, AD000046
• Expected Result:
  • Processing continues normally
  • Sequential tracking may work for AD000044 → AD000045
  • All images completed

Step 7: Validate Recovery Success

• Action: Check GPS estimates for all 4 images
• Expected Result:
  • AD000042: Accurate
  • AD000044: Recovered via L2/L3, error < 200m (AC-4), preferably < 50m
  • AD000045-046: Accurate

Step 8: Test Dataset B (Larger Gap)

• Action: Repeat test with AD000003, AD000009 (5-frame gap)
• Expected Result: Similar recovery, L2 successfully relocalizes

Success Criteria

Primary Criterion (AC-4):

• System recovers from zero-overlap scenarios
• Relocated image within < 200m of ground truth (AC-4 requirement)
• Processing continues without manual intervention

Supporting Criteria:

• L1 failure detected appropriately
• L2 retrieves correct region (top-5 accuracy)
• L3 refines to < 50m accuracy
• All images in sequence eventually processed

Expected Results

Dataset A:

Images: AD000042, AD000044, AD000045, AD000046
Gap: AD000043 skipped (simulates sharp turn)

Results:
- AD000042: L1 tracking, Error 21.3m ✓
- AD000043: SKIPPED (not in dataset)
- AD000044: L2 recovery, Error 38.7m ✓
  - L1 failed (overlap ~0%)
  - L2 top-1 retrieval: correct tile
  - L3 refined GPS
- AD000045: L1/L3, Error 19.2m ✓
- AD000046: L1/L3, Error 23.8m ✓

L1 Failure Detected: Yes (AD000042 → AD000044)
L2 Recovery Success: Yes
Images < 50m: 4/4 (100%)
Images < 200m: 4/4 (100%) per AC-4
AC-4 Status: PASS

Dataset B:

Images: AD000003, AD000009
Gap: 5 frames (AD000004-008 skipped)

Results:
- AD000003: Error 24.1m ✓
- AD000009: L2 recovery, Error 42.3m ✓

L2 Recovery Success: Yes
AC-4 Status: PASS

Pass/Fail Criteria

TEST PASSES IF:

• L1 failure detected when overlap < 5%
• L2 successfully retrieves correct region (top-5)
• Recovered image within 200m of ground truth (AC-4)
• Preferably < 50m (demonstrates high accuracy)
• Processing continues after recovery

TEST FAILS IF:

• L2 retrieves wrong region (relocalization fails)
• Recovered image > 200m error (violates AC-4)
• System halts processing after L1 failure
• Multiple recovery attempts fail

Analysis

Sharp Turn Characteristics (per AC-4):

• Next photo overlap < 5% or zero
• Distance < 200m (banking turn, not long-distance jump)
• Angle < 70° (heading change)

Recovery Pipeline:

1. L1 detects failure (low inlier ratio)
2. L2 (AnyLoc) global place recognition
3. L3 (LiteSAM) metric refinement
4. Factor graph incorporates GPS anchor

Why L2 Works:

• DINOv2 features capture semantic layout (roads, field patterns)
• VLAD aggregation creates robust place descriptor
• Faiss index enables fast retrieval
• Works despite view angle differences

Notes

• AC-4 specifies < 200m, but system targets < 50m for high quality
• Sharp turns common in wing-type UAV flight (banking maneuvers)
• L2 is critical component - if L2 fails, system requests user input per AC-6
• Test validates "kidnapped robot" problem solution