MIT-Media-Lab/hot3d-vitra-atomic-instructions-v1

HOT3D VITRA Atomic Instructions v1

This is the language-conditioned, atomic-episode derivative of MIT-Media-Lab/hot3d-vitra-strict-v1. It contains 32,850 hand-specific episodes, references 250 unique strict source videos, and indexes 1,227,851 target-hand frames.

Construction

• Source: immutable strict-v1 commit 191a139da6dec2e55f73c1558679fd791b19e473.
• Left and right hands are segmented independently inside continuous aligned MANO runs.
• Recording-world wrist trajectories are Gaussian-smoothed (sigma 2) and cut at speed minima in a 25-frame window. Boundaries receive up to 3 context frames on each side.
• Clips require at least 15 frames and at least 50% target-hand strict frames.
• Eight strict frames with an in-image six-joint palm marker are supplied to Qwen/Qwen3.5-122B-A10B-FP8 using the hand-specific VITRA JSON prompt.
• Parse failures and N/A captions are quarantined and excluded.

Object names are retained as audit metadata only and are not supplied to the caption model. Captions are imperative, one-sentence, target-hand-specific instructions. This v1 release contains the primary caption only; rephrases are left empty.

Coordinates and representations

Every atomic .npy preserves its source recording's HOT3D MoCap world frame, per-frame World2Cam camera extrinsics, MANO pose, and 21 keypoints. No pose, joint, or camera value is smoothed, re-forwarded, or rewritten during atomic segmentation/materialization. transl_worldspace is MANO wrist joint 0 in metres. Training chunks are transformed to the observation/anchor camera in the VITRA dataloader, not persisted in pre-anchored coordinates.

Both MANO-angle and 21-keypoint normalization statistics are included.

Layout

Annotation/hot3d_atomic_train/episode_frame_index.npz
Annotation/hot3d_atomic_train/episodic_annotations/*.npy
Annotation/statistics/hot3d_atomic_train_angle_statistics.json
Annotation/statistics/hot3d_atomic_train_keypoints_statistics.json
Video/HOT3D_root/{Aria,Quest3}/<recording>/<stream>.mp4
manifests/accepted.jsonl
manifests/quarantine.jsonl
manifests/candidate_manifest.jsonl
manifests/caption_ledger.jsonl
build_report.json
validation_report.json

The annotation dictionaries require numpy.load(..., allow_pickle=True) in a trusted environment. video_decode_frame contains absolute frame indices in the shared source recording MP4. Text intervals are half-open [start,end).

Validation

The published build validator status is pass. Maximum world-to-camera joint consistency error is 1.093e-07 m; there are 0 schema/geometry failures, no missing referenced videos, and no unreferenced uploaded videos.

License and attribution

This is a derivative of Meta's HOT3D dataset. HOT3D sequence data and hand annotations have distinct share-alike/noncommercial terms; users must follow all applicable official HOT3D licenses and attribution requirements. MANO model assets are not redistributed and remain subject to their own license.