play_human_fruits_v2_train_eef

Human EEF annotations for aytsaiusc/play_human_fruits_train.

2026-08-26: rebuilt after fixing a hand-detection defect. If you pulled this dataset before that date, re-pull — the earlier build annotated only 87.6% of frames (worst episode 50.9%) and its hand boxes were systematically cut off.

What was wrong with the earlier build

Hand detection ran Grounding-DINO on a hard-coded workspace window whose right edge sat at x=440. The hand enters this rig from the right, so the window cut it in half: the detector only ever saw the fingertips and boxed those. HaMeR then estimated the pose from a fingertip-only crop, so the 21 keypoints and everything derived from them were wrong. A downstream shape filter (aspect < 1.8) then rejected many of those boxes, dropping the whole frame.

Net effect on the previous release: 59.3% of boxes were cut off at the window edge, and only 87.6% of frames had any annotation at all (worst episode: 50.9%).

The fix (both parts are required)

1. prompt "a human hand" (was "a hand") and detect on the full frame, no window. "a hand" boxes the whole arm on top-down footage (median box 302 px wide, aspect 2.97); the window existed only to chop the forearm off, and chopped the hand too.
2. shape filter aspect < 2.2 (was 1.8). Correct hand boxes are naturally a little elongated from above (median aspect 1.75), so 1.8 rejected most of them. Changing only the prompt makes things worse — measured: detection drops 87.7% → 81.8%.

Verified over all 300 episodes / 219087 frames

	previous	this release
frames with annotation, mean / median	87.6% / 89.7%	99.9% / 100.0%
worst episode	50.9%	94.4%
per-episode change	—	272 better / 2 worse, median +10.3pp
median box width	109 px	135 px
boxes cut off at the window edge	59.3%	0.8%
lift_ok (depth lift converged), mean	87.5%	99.9%

★ What this fix does and does not buy you

It buys coverage, not precision. On frames that both releases annotate, accuracy is unchanged: measured against an independent signal (HSV fruit-mask centroid, nothing to do with hand detection), the pinch centre sits 12.6 px from the carried fruit in the old release and 13.1 px in this one — the same within noise. Per-frame jitter is marginally lower (0.99 → 0.93 px).

That is self-consistent: the old shape filter discarded frames whose box was wrong, so what survived into the old release was already the well-boxed subset. What you gain here is the ~12% of frames that used to be thrown away entirely — concentrated in the reach / pre-contact phase and parts of the carry phase.

Do not read the keypoint-referenced numbers (pinch vs thumb/index midpoint, 6.4 → 5.3 px) as evidence of better accuracy: the eef is derived from those same keypoints, so that comparison is circular.

Contents (per frame)

column	shape	meaning
observation.eef.position_right	(3,)	pinch center, world frame (metric, depth-lifted)
observation.eef.rotation_matrix_right	(9,)	EE orientation, world frame (smoothed)
observation.eef.width_right	(1,)	grip opening — see caveat
observation.eef.detected_right	(1,)	hand detected
observation.eef.lift_ok_right	(1,)	depth lift converged
observation.eef.kpts21_2d	(21,2)	HaMeR keypoints, full-frame pixels
observation.eef.kpts21_3d_world	(21,3)	HaMeR keypoints, world frame (metric)
observation.eef.kpt2d_crop128_*	(2,)	see caveat
observation.eef.{position,width}_right_raw		pre-smoothing

The eef triple used downstream is [pinch, p + R·w/2, p − R·w/2] — all three sit at the grasping end; there is deliberately no wrist point (the robot-side triple is built the same way, so the two domains match).

Caveats

Episode 264 is absent (299 of 300). The source dataset is internally inconsistent there: its mp4 has 1222 frames while its parquet has 567 rows (neighbours 263/265 match exactly). Those 567 depth rows cannot be aligned to a known segment of the video, so the episode is skipped rather than silently mis-aligned. Episode indices are kept identical to the source (0..263, 265..299), so there is a gap; meta/ describes the 299 present, and the original 300-episode metadata is kept alongside as *.orig300. Episode 264's videos remain under videos/, without a parquet or a meta/ entry.

width_right is not a reliable magnitude. Against the HaMeR thumb-tip/index-tip separation over 1612 frames: p5 0.63, median 0.85, p95 1.16, full range 0.27–3.69, and 22% of frames exceed 1.0. It is not a constant rescaling. Use the two finger points for where the grasp is; do not treat the opening as a metric distance, and do not expect it to indicate whether the hand is holding something.

kpt2d_crop128_* are stale. They were computed with a crop belonging to a different dataset and are wrong here. Use kpts21_2d (full-frame) with your own crop, or the 3D columns.

Episode 230 is the one episode that got worse (100% → 94.4% annotated). Not investigated.

Provenance

AYTSAI_VERSION=fruits_v2_fix throughout; phantom Pass1 re-run with PHANTOM_BBOX_FIX=1 (prompt + full frame + PHANTOM_ASPECT_MAX=2.2). The previous release was built from the same source videos and is left untouched for comparison.