Phase 22 · BaselineLoading results
Frame image not loaded. Select its local source image to view measured boxes in context.![]()
Phase 25 · Failure Atlas
See where the model fails.
86 verified frames. Six camera conditions. Compare measured original-model predictions.
Frame explorer
Find the failure.
Both original models are measured across 516 paired views. The 860 within-model clean-vs-stressed pairs are a separate comparison; none of these views is independent validation.
FRAME 001 / 086
Loading the verified 86-frame image set across all six reconstructed conditions… Bundled images are rebuilt from the checksum-verified KIOS archive. Images selected locally are not hash-checked. Dataset source ↗ · Method ↗
Ground truthMatched detectionUnmatched detection
Phase 23 · Original robust detectorLoading measured results
Measured frame predictions
Published condition aggregate · Clean: mAP50 — · recall —. These aggregates are distinct from the selected frame metrics above.
Original-checkpoint replay: all 24 aggregate cells and four prediction tables matched ↗Filter the record
Where does it break?
Descriptive filters only. These frames have already been evaluated.
Loading viewsSelect a cell to inspect
Target matched Target missed Filtered out
Confidence and correctness
Confident ≠ correct.
IoU with nearest target
True positive False positive
Hover a point or select a case to inspect its score and IoU.
Detector score →
Showing measured baseline boxes with score ≥ 0.01, plus any matched boxes below it. All 136,359 predictions are available in the complete table. Box correctness does not measure landing safety.
Evidence and provenance
Show the work.
Inputs verified. Baseline and Phase 23 predictions evaluated on 86 protected test frames. 5,160 treatment views analyzed in controlled occlusion topology study.