LeoJiangOR/hot3d-clips-vitra-streaming-v2

INCOMPLETE — SUPERSEDED — DO NOT TRAIN

This interrupted development release used MANO surface-to-box IoU as a hard camera-alignment gate. HOT3D's official amodal hand boxes follow the UmeTrack surface, and participant-specific UmeTrack/MANO surface differences therefore caused false quarantines even when their landmarks and RGB overlays agreed. Processing was stopped before COMPLETE. Use LeoJiangOR/hot3d-clips-vitra-streaming-v3, which separates UmeTrack-to-box camera validation from UmeTrack-to-MANO landmark validation.

HOT3D-Clips VITRA Streaming v2 (superseded)

This repository is being populated. Do not train from it until COMPLETE exists.

This is a VITRA-ready derivative of the 2,804 official HOT3D-Clips training clips at source revision 30fe9674782f32e1e5edba98476b6ff4300132c5. It is not the complete 833-minute raw HOT3D release. HOT3D-Clips contains non-overlapping, curated five-second clips at 30 Hz. Official test clips are excluded because they do not publish hand/object ground truth.

What is retained

• Aria: the egocentric RGB stream 214-1.
• Quest 3: both egocentric monochrome streams 1201-1 and 1201-2, encoded as three-channel video for the current VITRA loader.
• Per-frame MANO rotations, beta, canonical wrist joint, 21 joints in world and camera coordinates, object transforms, visibility masks, exact timestamps, and world-to-camera transforms.
• A deterministic post-hoc weak instruction based on full-clip object motion, hand motion, and hand-object proximity. No LLM caption is used as target text; the use of future trajectory metrics is explicitly recorded.

HOT3D cameras use FISHEYE624, while the current VITRA loader accepts a 3x3 pinhole matrix. Images are therefore warped with Meta's official hand_tracking_toolkit camera implementation into a 640x640 upright pinhole view. The stored intrinsic and per-frame extrinsic describe that exact output view. Aria's native sideways storage is rotated clockwise by 90 degrees.

HOT3D's six-dimensional mano_pose.wrist_xform translation is the SMPL-X MANO model origin, not the canonical MANO wrist joint. The released VITRA transl is joint 0 obtained by forwarding MANO; the original model-origin translation and 15-D PCA pose are preserved as provenance sidecars. MANO model files are never redistributed.

Quality and split policy

Every source tar is pinned by LFS SHA-256 and checked against the exact official WebDataset schema. Every output video is decoded for all 150 frames. MANO is cross-checked against independently supplied UmeTrack landmarks; projections, visibility, timestamps, rotations, VITRA keypoint/angle loader paths, and exact statistics accumulators are audited per clip. A camera with too little valid visible hand supervision is rejected without discarding another valid camera from the same clip. A selected camera with duplicated/dropped (non-30Hz) source timestamps is quarantined rather than silently treated as a uniform trajectory. Every uploaded file is size-verified at an immutable commit, and the per-clip manifest is downloaded and hash-verified before local cleanup.

The derived validation split holds out participants P0014 (Aria) and P0017 (Quest 3); all other official-training participants are train. sequence_id is the leakage/sampling group, so clips from one recording must not be split again.

License and attribution

The source sequence data is licensed CC BY-SA 4.0 and the hand annotations are licensed CC BY-NC-SA 4.0. Because this derivative includes the hand annotations, users must satisfy the more restrictive noncommercial/share-alike terms as well as HOT3D attribution requirements. Object models are not redistributed here. See the official HOT3D repository, the official HOT3D-Clips source card, and the included source license provenance. Cite the HOT3D paper when using this release.

Limitations

The commands are post-hoc metric-derived weak supervision, not recorded human language; using them as conditioning can leak coarse action identity from the full trajectory. Quest 3 images are grayscale. Five-second curated clips are useful for hand action learning but do not reproduce the long-horizon distribution of full recordings. This dataset is egocentric human-hand data, not robot proprioception.