LeoJiangOR/dexycb-vitra-streaming-v2

DexYCB VITRA Hybrid Streaming Release

This is a completed, audited VITRA Stage-1 conversion of all 1,000 physical DexYCB sequences. The release contains 1000 usable physical sequences and 0 quarantined sequences. COMPLETE is present only after every source sequence reached a terminal state, all global indices/statistics were rebuilt, and the final commit was checked by immutable revision.

Exact source policy

The source choice is made once per subject; there is no fuzzy filename match, camera fallback, nearest-frame lookup, or per-file mixing.

subject	qualified component	source archive SHA-256
20200709-subject-01	official-complete	ce44e97a7567f7fb128b474ae33f327715314956a1c7b24cb08b08bc51cf4303
20200813-subject-02	qualified-world-shard	a55757de56ca941bbf407d0989062f034cecde752eaf5ed6c7e06b12dcdbbec6
20200820-subject-03	qualified-world-shard	48323cb3982bf481adf0e3c110615724ccaf32100307ce02cb1d0df6fa41f8fb
20200903-subject-04	qualified-world-shard	85da9628de98e72d2df74c18038e74328ae6288d0a90cf9286f3dfaa98516cb7
20200908-subject-05	qualified-world-shard	6e486214a381a05590f14a2b56297e64a3065e1f8b567e8d57df6aac44e160e5
20200918-subject-06	qualified-world-shard	80ceb2b3ee5c575ea02b259b3f765d6b14b96b9f79f5f3e6b6bfb4144b564293
20200928-subject-07	official-complete	af34bbf19bf02f695c9680fefa76362ad8ffe3507bf3237425fb5851db492522
20201002-subject-08	official-complete	f65201fc619b6aa0d6053d16a372c6776573cf6db476c59e3f69c2c56ed4f480
20201015-subject-09	official-complete	ceecf8c9db6768db2dd1c24320915415b6c9a4bb6db6ae7937012e832a7e8bdc
20201022-subject-10	qualified-world-shard	2236c553f2ae71246d73af22b92e35fb6f4e0e4677b0f6c44aee510f1290c524

Subjects 01, 07, 08, and 09 use the complete official archives. Subjects 02--06 and 10 use a hash-pinned RGB archive plus qualified world-space MANO, joints, and target-object shards. For the latter subjects, aligned depth, official segmentation, non-target object trajectories, and camera-space MANO parameters are explicitly unavailable and were not synthesized. Their full qualification reports and normalized 1,000-sequence source index are published under Audit/source_provenance/.

The 600 recovered sequences deliberately retain two non-interchangeable evidence tiers. Exactly 580 are redundant-camera-closure: independent HandOccNet camera rows close against the published world labels and every sequence has a passed reference adjudication. The remaining 20 subject-02 S0-validation sequences are explicitly single-derived-source: no independent HandOccNet camera rows exist for that split, so no reference adjudication is invented; only the hash-pinned published world source and its internal world-to-joint/root and target-mesh closure checks are available. The independent release verifier requires this exact 580/20 partition. Consumers that require redundant camera evidence can filter on source.world_labels.reference_evidence.evidence_tier in each sequence manifest.

Training organization

All eight cameras of a physical action remain in the same official S0 split. Low-observability cameras may be rejected independently, while the physical sequence remains usable through its accepted views. Equal-action sampling metadata is available under Build/sampling_groups/ to prevent an action with more accepted views from receiving more weight.

split	accepted camera episodes	indexed anchors
train	6260	382790
val	318	21462
test	1260	74800

Files and VITRA loader contract

RGB episodes are under Video/DexYCB_root/. Labels are NumPy dictionaries at Annotation/dexycb_<split>/episodic_annotations/<episode_id>.npy. Each label contains the exact video_name, video_decode_frame, source frame IDs and timestamps, calibrated intrinsics, static world_to_camera extrinsics, world/camera 21-joint arrays, MANO beta/rotation-matrix pose/translation fields, the target-object trajectory, causal task text, and training_anchor_valid.

The final Annotation/dexycb_<split>/episode_frame_index.npz contains index_frame_pair and index_to_episode_id. Column 0 of each pair selects an episode ID; column 1 selects a label row. Only rows whose training_anchor_valid value is true are indexed. A minimal lookup is:

import numpy as np

with np.load(index_path, allow_pickle=True) as index:
    episode_slot, frame_id = index["index_frame_pair"][sample_id]
    episode_id = index["index_to_episode_id"][episode_slot]
episode = np.load(
    label_dir / (episode_id + ".npy"), allow_pickle=True
).item()
rgb_frame_id = int(episode["video_decode_frame"][frame_id])
assert bool(episode["training_anchor_valid"][frame_id])

Both keypoints and MANO angle statistics are exact merges of per-sequence float64 accumulators. DexYCB is fixed-camera, allocentric, short-horizon human pick-up data. It is useful supplemental VITRA supervision, but it is not a standalone egocentric or robot-control corpus. Its deterministic language is derived from the official pick-up protocol and object/hand metadata; the future-observing visual-language audit is never used as training text.

Reproducibility and audit

Every sequence manifest declares all payload SHA-256 values, immutable upload commit, source provenance, pipeline version, and conversion-code hashes. The authoritative merged SQLite ledger is Build/release_ledger.sqlite3; the final file inventory is Build/final_manifest.json. An independently regenerated release audit is uploaded to Audit/independent_release_verification.json. That audit materializes the immutable per-sequence report bytes, checks their manifest SHA-256 and size, and requires all nine wrist/joint temporal metrics in each report to exactly match the corresponding release-ledger row before it recomputes the global temporal exclusion policy.

Please cite DexYCB (CVPR 2021), follow its CC BY-NC 4.0 license, and separately obtain MANO under its own license.