GigaHands strict-v1 full quality audit
Coverage
| audit | coverage |
|---|---|
| converted episodes | 12,783 / 12,783 |
| structural/temporal label frames | 3,227,699 / 3,227,699 |
| RGB projection-detector frames | 63,915 / 63,915 |
| RGB policy | every episode, up to 5 evenly spaced frames |
Results
| category | episodes | rate |
|---|---|---|
| structural fatal error | 0 | 0.000% |
| cam0: neither hand visible for more than half the episode | 1,353 | 10.584% |
| cam0: neither hand visible for the whole episode | 399 | 3.121% |
| persistent MANO-vs-triangulation disagreement (>5 cm median) | 6 | 0.047% |
| output wrist temporal jump (>25 cm/frame) | 2 | 0.016% |
| hand points behind selected camera (>5% joints) | 31 | 0.243% |
| RGB detector: automatic alignment suspect | 18 | 0.141% |
| RGB suspect: manually confirmed mismatch | 0 | 0.000% |
| RGB suspect: unresolved after review | 0 | 0.000% |
| RGB decode incomplete | 0 | 0.000% |
Neither hand projects into cam0 for 427,709 / 3,227,699 label frames (13.251%). These are not necessarily corrupt GigaHands motions, but they provide little or no image evidence for camera-conditioned VITRA training.
The detector raised 18 automatic RGB-alignment suspects. Manual review cleared
18: in those clips the projected MANO hand follows the real hand,
while the detector fires on a monitor, object, toy dog, or a partially cropped hand with the wrong
handedness. There are 0 confirmed mismatches and
0 unresolved automatic suspects. The review decision file is
rgb_alignment_manual_review.json, and the rendered evidence is in the rgb_*_review_*.jpg files.
There are 213 episodes with a large MANO/triangulation tail discrepancy.
In 185 of them the triangulated reference has a >10 cm jump and
is more than twice as discontinuous as the converted MANO output, so those are more consistent
with a noisy independent reference than with a converted-label failure. The remaining
28 are not automatically attributable to either source and stay
in the conservative exclusion/review list. Their maximum-disagreement image overlays are in
tail_disagreement_review_*.jpg.
Conservative training recommendation
Excluding low-visibility cam0 episodes, persistent or unresolved geometry disagreements, output
temporal jumps, detector-confirmed RGB mismatches, and decode failures removes 1,384
episodes (10.827%). The exact IDs are in
recommended_exclude_episode_ids.json; all category measurements are in audit_categories.json.
This is deliberately conservative. In particular, the visibility filter measures usefulness for VITRA training, not corruption of the underlying multiview motion. A better recovery policy is to select another official camera with high projected-hand visibility when that video is available, and only drop an episode when no camera is suitable.