LeoJiangOR/gigahands-vitra-mano-strict-v1

GigaHands → VITRA official-MANO strict-v1

VITRA Stage-1 annotations derived from GigaHands' official MANO fit, regenerated with strict video/annotation binding and accompanied by an exhaustive numerical audit plus an every-episode RGB audit. This repository contains annotations only: it does not redistribute GigaHands videos or the licensed MANO model.

item	value
source manifest entries	13,249
converted episodes	12,783
train / test episodes	11,461 / 1,322
converted frames	3,227,699
train / test frames	2,963,261 / 264,438
camera	brics-odroid-001_cam0
action interface	VITRA keypoints mode (16 × 192 future action chunks at training time)

Although the source is the official MANO fit, this is not a MANO-angle-only target dataset. Each episode contains MANO-derived 3-D joints, wrist transforms, MANO shape, global orientation, and finger rotation matrices. Current VITRA training uses the 3-D joint/keypoint action interface.

What changed from yuzhench/gigahands-vitra-mano

The older release retained 453 known video-length mismatch candidates. Strict-v1 applies the following deterministic gates before writing an episode:

1. bind (scene, sequence, camera) to the exact video in GigaHands' multiview_camera_video_map.csv, never to a sorted file-list position;
2. require exact temporal agreement among left/right MANO parameters and keypoints_3d_mano;
3. reject video/annotation length differences above 12 frames and trim only a short video tail within that tolerance;
4. require at least 0.9 raw-triangulation validity using chosen_frames_{left,right}.json, rather than treating the always-finite MANO-filled output as 100% valid.

Conversion removed 464 entries: 381 for video/annotation length disagreement and 83 for missing official video mapping. One entry lacked MANO keypoints and one was shorter than 32 frames. None of the older release's 453 known length-misaligned episode IDs remains in strict-v1.

Reproduction code is in kaichen-z/VLA-HAND@4458502, including:

• scripts/build_gigahands_official_mano_strict_v1.sh
• scripts/run_gigahands_full_quality_audit.sh
• data/preprocessing/convert_gigahands_to_vitra_stage1.py
• tools/audit_gigahands_full.py

Contents

Annotation/
├── gigahands_real_train/
│   ├── episodic_annotations.tar.zst       # 11,461 .npy episodes
│   ├── episode_frame_index.npz            # 2,963,261 frame/window-start entries
│   └── conversion_report.json
├── gigahands_real_test/
│   ├── episodic_annotations.tar.zst       # 1,322 .npy episodes
│   ├── episode_frame_index.npz            # 264,438 entries
│   └── conversion_report.json
└── statistics/
    ├── gigahands_real_train_keypoints_statistics.json
    └── gigahands_real_test_keypoints_statistics.json
subset_manifest.json                        # original 13,249-entry source manifest
conversion_report.json
quality_rejections.json
quality_comparison.{json,md}
baseline_residual_misaligned_episodes.json
artifact_manifest.json                       # byte size and SHA-256 for every uploaded artifact
quality/full_quality_audit/
├── full_audit_report.{json,md}
├── recommended_exclude_episode_ids.json
├── audit_categories.json
├── labels_summary.json
├── labels_per_episode.jsonl
├── video_summary.json
├── video_per_episode.jsonl
├── video_samples.jsonl
├── rgb_alignment_manual_review.json
└── *_review_*.jpg

The .tar.zst files contain an episodic_annotations/ directory. Extract them in place:

huggingface-cli download \
  --repo-type dataset \
  LeoJiangOR/gigahands-vitra-mano-strict-v1 \
  --local-dir datasets/vitra_gigahands_official_mano_strict_v1

for split in train test; do
  tar --zstd -xf \
    datasets/vitra_gigahands_official_mano_strict_v1/Annotation/gigahands_real_${split}/episodic_annotations.tar.zst \
    -C datasets/vitra_gigahands_official_mano_strict_v1/Annotation/gigahands_real_${split}
done

mkdir -p datasets/vitra_gigahands_official_mano_strict_v1/Video
ln -s /path/to/GigaHands/multiview_rgb_vids \
  datasets/vitra_gigahands_official_mano_strict_v1/Video/GigaHands_root

Episode schema

Each .npy loads as a Python dictionary. left and right contain:

field	shape	meaning
beta	(10,)	official MANO shape
global_orient_worldspace	(T,3,3)	wrist/global orientation in the GigaHands world frame
global_orient_camspace	(T,3,3)	orientation in the selected camera frame
hand_pose	(T,15,3,3)	MANO finger rotations with the MANO mean pose restored
transl_worldspace	(T,3)	wrist position, equal to joints_worldspace[:,0]
transl_camspace	(T,3)	wrist position in camera coordinates
joints_worldspace	(T,21,3)	official-MANO joints with global translation restored
joints_camspace	(T,21,3)	joints in camera coordinates
kept_frames	(T,)	validity mask

Episodes also contain video_name, video_decode_frame, language instruction, intrinsics, extrinsics, and camera metadata. Sixteen-frame action windows are created by the VITRA loader; each stored .npy is a complete episode rather than one training window.

Full quality audit

Coverage:

• 12,783 / 12,783 episodes;
• 3,227,699 / 3,227,699 label frames for schema, finite values, rotations, world/camera transforms, temporal continuity, distortion-aware camera projection, and comparison with independent raw triangulated keypoints;
• 63,915 / 63,915 RGB checks: five evenly spaced frames in every episode, using a hand-pose detector and manual review of every automatic mismatch candidate.

Main results:

category	episodes	rate
structural fatal error	0	0.000%
likely left/right identity swap	0	0.000%
neither hand visible in cam0 for more than half the episode	1,353	10.584%
neither hand visible in cam0 for the entire episode	399	3.121%
persistent MANO/triangulation disagreement above 5 cm median	6	0.047%
converted-output wrist jump above 25 cm/frame	2	0.016%
automatic RGB alignment suspects	18	0.141%
manually confirmed RGB mismatch among those suspects	0	0.000%

The typical per-episode MANO-versus-independent-triangulation median MPJPE is approximately 7 mm. Of 213 tail-disagreement candidates, 185 are more consistent with a discontinuous triangulation reference; the remaining 28 are retained in the conservative exclusion/review category.

quality/full_quality_audit/recommended_exclude_episode_ids.json contains a conservative episode-level training policy: 1,384 episodes (10.827%). Applying it leaves 11,399 episodes and 2,863,827 frames (train: 10,296 episodes / 2,637,406 frames; test: 1,103 episodes / 226,421 frames). Most exclusions are not proven source corruption: they are cam0 visibility failures. If another official camera is available, selecting the camera with the best projected-hand visibility is better than discarding the underlying multiview motion. If training must remain cam0-only, use this list or, preferably, filter individual 16-frame windows by hand visibility.

The 18 automatic RGB suspects were all manually cleared. Typical detector failures were firing on a monitor, an object, or a toy dog while the projected MANO joints still followed the real, partially cropped hand. Review images and the decisions are included instead of silently converting detector output into deletion labels.

Video mapping

Videos are not included. Obtain them under the original GigaHands terms and place or link multiview_rgb_vids at Video/GigaHands_root. video_name in every episode and video_path/start_frame/end_frame in subset_manifest.json give the exact source mapping.

Do not infer a sequence's video by list index: sequence IDs contain gaps. Strict-v1 verifies the official CSV mapping during conversion.

Licence and citation

These are derived GigaHands annotations. The original images, videos, MANO fits, and their usage terms remain with the GigaHands authors. Follow the original GigaHands licence and citation requirements. The MANO model is not redistributed and must be obtained separately under its own licence. The annotation format follows Microsoft VITRA Stage-1.