gps-denied-onboard/docs/03_tests/37_sharp_turn_minimal_overlap_spec.md

Acceptance Test: Sharp Turns with Minimal Overlap (<5%)

Summary

Validate AC-4 requirement for handling sharp turns where consecutive images have minimal overlap (1-5%) but do share some common features, testing the boundaries of L1 sequential tracking before L2 activation.

Linked Acceptance Criteria

AC-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%.

Preconditions

• ASTRAL-Next system operational
• LightGlue configured with adaptive depth mechanism
• L2 global place recognition ready for fallback
• Minimal overlap test dataset prepared

Test Data

• Dataset A: AD000032-AD000035 (contains 220.6m jump)
• Dataset B: AD000042, AD000044, AD000045 (skip 043, natural gap)
• Overlap Characteristics: 2-5% estimated (edges of consecutive frames)
• Distance: All jumps <230m (within 200m drift requirement)

Test Steps

Step 1: Test Minimal Overlap Scenario A

Action: Process AD000032 → AD000033 (220.6m jump) Expected Result:

• L1 SuperPoint+LightGlue attempts matching
• Feature count low (10-50 matches vs normal 100+)
• LightGlue adaptive depth increases (more attention layers)
• Confidence score reported as LOW
• Status: MINIMAL_OVERLAP_L1

Step 2: Verify L1 Handling or L2 Activation

Action: Analyze matching result and system response Expected Result:

• IF ≥10 high-quality matches: L1 succeeds, pose estimated
• IF <10 matches: L1 fails gracefully, L2 activates
• No system crash or undefined behavior
• Transition between L1/L2 smooth
• Status: L1_OR_L2_SUCCESS

Step 3: Test Minimal Overlap Scenario B

Action: Process AD000042 → AD000044 (skipping 043) Expected Result:

• Larger gap increases difficulty
• L1 likely fails (overlap too low)
• L2 AnyLoc retrieval activates
• DINOv2 features match despite view change
• Global position estimated from satellite database
• Status: L2_RELOCALIZATION

Step 4: Validate Position Accuracy

Action: Compare estimated positions with ground truth Expected Result:

AD000033: Error <100m (minimal overlap, higher uncertainty)
AD000044: Error <100m (L2 retrieval, coarse localization)
Drift: <200m for both scenarios (AC-4 requirement met)

Step 5: Test Continuous Minimal Overlap Sequence

Action: Process AD000032-035 as continuous sequence Expected Result:

• Initial minimal overlap (032→033) handled
• Subsequent frames (033→034→035) easier (normal overlap)
• Factor graph smooths trajectory
• Final accuracy improved through optimization
• All frames: Error <50m after optimization
• Status: SEQUENCE_COMPLETED

Step 6: Verify Angular Constraint (<70°)

Action: Analyze estimated rotation between frames Expected Result:

• Rotation estimates <70° for all minimal overlap scenarios
• Rotation estimates consistent with wing-type UAV banking
• No unrealistic 180° rotations estimated
• Status: ANGULAR_CONSTRAINT_MET

Pass/Fail Criteria

PASS if:

• Both minimal overlap scenarios processed without crash
• Position drift <200m for each minimal overlap frame
• Angular changes <70° estimated
• System demonstrates graceful L1→L2 transition
• Final optimized trajectory <50m error

FAIL if:

• System crashes on minimal overlap
• Position drift >200m for any frame
• Angular changes >70° estimated
• No L2 fallback when L1 fails
• Trajectory diverges

Technical Validation

• Match Count: Minimal overlap yields 10-50 matches (vs normal 100+)
• LightGlue Depth: Adaptive mechanism uses more layers (8+ vs typical 6)
• L1 Confidence: <0.5 for minimal overlap (vs >0.8 normal)
• L2 Activation: Triggered when L1 confidence <0.4
• Retrieval Accuracy: Top-1 satellite tile correct in ≥90% cases

Edge Case Analysis

• 1-5% overlap: Boundary between L1 possible and L1 impossible
• 5-10% overlap: L1 should succeed with high confidence
• 0% overlap: Covered in separate test (36_sharp_turn_zero_overlap_spec.md)

Notes

• This test validates the smooth transition zone between L1 and L2
• LightGlue's attention mechanism is crucial for minimal overlap success
• Real flights may have brief periods of minimal overlap during banking
• AC-4 allows up to 200m drift, giving system reasonable tolerance