MIT-Media-Lab/hot3d-vitra-strict-v1

HOT3D VITRA Strict v1

This dataset is the full-recording VITRA-ready HOT3D derivative produced by the hot3d-strict-v1 cleaning pipeline. It is separate from the curated five-second hot3d-clips-vitra-streaming-v3 release.

Only recordings that pass the strict quality gate are included. The release contains 250 selected egocentric videos and matching episodic annotations:

• 126 Aria videos from stream 214-1;
• 77 Quest 3 videos from stream 1201-1;
• 47 Quest 3 videos from stream 1201-2;
• 889,553 decoded video frames;
• 792,056 strict-valid training anchors.

For Quest 3, one camera view is selected per recording using the higher strict-valid-frame ratio. The two views are not treated as two independent action samples.

Strict quality gate

A frame is strict-valid when at least one hand:

1. has aligned and finite MANO/keypoint data;
2. has valid MANO rotations and consistent world/camera geometry;
3. passes camera-pose, HOT3D QA, and exposure masks;
4. is not marked as a short temporal outlier; and
5. has at least 10 of 21 projected joints in front of the camera and inside the image.

A recording is retained when at least 50% of its frames are strict-valid. From 294 source recordings, 250 pass and 44 are excluded. All 44 exclusions have reason low_hot3d_valid_ratio. All 250 retained episodes pass the final structural, video, and geometry validator.

The quality pipeline does not use triangulated validity. It does not smooth or rewrite MANO values; it preserves source-derived MANO/joints and applies strict kept-frame masks.

Annotation contents

Each .npy file is a pickled Python dictionary containing:

• video-relative frame indices and source timestamps;
• a per-frame World2Cam camera extrinsic and calibrated intrinsic;
• left/right MANO beta, world/camera global orientation, and 15 local joint rotation matrices;
• left/right wrist translation and 21 joints in recording-world and camera coordinates, in metres;
• projected 2D keypoints and visible-joint counts;
• source/aligned/interpolated/strict-kept masks and provenance;
• headset, stream, participant, objects, camera convention, and quality-gate metadata.

The canonical stored hand trajectory is in the HOT3D MoCap recording world. Each recording has its own world frame. For VLA training, a future action chunk must be transformed into the observation/anchor camera coordinate system in the dataloader; the dataset does not persist a separate coordinate system per training chunk.

transl_worldspace and transl_camspace are canonical MANO wrist joint 0, not the MANO model origin. The original MANO root translation and source PCA pose are retained as provenance. MANO model files are not redistributed.

Language status

HOT3D does not provide aligned natural-language action instructions for these full recordings. text.left and text.right are empty, and source.instruction_status is pending_atomic_segmentation_and_per_hand_captioning. Do not interpret object names as action instructions.

Layout

Annotation/hot3d_train/episode_frame_index.npz
Annotation/hot3d_train/episodic_annotations/*.npy
Video/HOT3D_root/{Aria,Quest3}/<recording>/<stream>.mp4
quality_sidecars/*.npz
strict_manifest.csv
strict_build_report.json
validation_report.json
quality_rejections.json
view_selection.csv
HOT3D_STRICT_DATA_REPORT.md

strict_manifest.csv is the authoritative list of uploaded training videos. view_selection.csv records the complete camera-view audit. Rejected raw or Stage1 videos are not included in this repository.

Loader safety

Annotations use numpy.save with nested dictionaries and therefore require numpy.load(path, allow_pickle=True) in a trusted environment. Use strict_frame_valid_any_hand and the per-hand kept_frames masks for training loss/index construction.

License and attribution

This is a derivative of Meta's HOT3D dataset. HOT3D source sequence data and hand annotations have different share-alike/noncommercial terms; because this release contains hand annotations, users must follow all applicable HOT3D license and attribution requirements. See the official HOT3D release for the authoritative license text and cite the HOT3D paper when using this data.

MANO model assets are not included and remain subject to their separate license.