Datasets:

MIT-Media-Lab
/

oakink2-vitra-streaming-v1

license: cc-by-sa-4.0
task_categories:
  - video-classification
  - image-to-3d
language:
  - en
tags:
  - oakink2
  - vitra
  - mano
  - human-object-interaction

OakInk2 → VITRA Stage-1 (complete audited release)

This repository contains all 627 physical OakInk2-TaMF sequences converted to VITRA Stage-1. Every sequence source pair is pinned to kelvin34501/OakInk-v2 revision 21705616140d726607027e70d58b7837f442ffd8, aligned by exact frame identity, converted across the four calibrated views, checked by geometry and every-frame RGB audits, smoke-tested through the VITRA loader, uploaded, and verified at an immutable commit before local cleanup.

split	physical sequences	accepted camera episodes	indexed anchors
train	431	1722	2368048
val	39	155	183925
test	157	627	805606

Files and VITRA loader contract

RGB episodes are under Video/OakInk2_root/. Labels are NumPy dictionaries at Annotation/oakink2_<split>/episodic_annotations/<episode_id>.npy. Each label contains the exact video_name, video_decode_frame, source frame IDs and timestamps, calibrated intrinsics, per-frame world_to_camera extrinsics, world/camera 21-joint arrays, MANO beta/rotation-matrix pose/translation fields, object trajectories, causal task text, and training_anchor_valid.

The final Annotation/oakink2_<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])

Before the final commit, the release gate independently re-reads all 2504 accepted camera reports and calibrates wrist/joint speed distributions over all 627 physical sequences. Publication is blocked if a maximum exceeds both its conservative absolute floor and ten robust deviations in log1p space. This release has 0 blocked temporal sequences; the complete evidence is Audit/temporal_quality.json.

The legacy RGB audit helper only distinguished train/test. Before final publication, all 39 official-val RGB summaries and every per-frame JSONL record are therefore normalized to val, their sequence-manifest hashes are updated, and every changed remote byte is downloaded and verified. This correction does not alter videos, MANO/keypoint labels, action chunks, or training indices.

The official OakInk2-TaMF split is pinned at commit 4cb9c461b92c032715374bd252c7179c70a14bde. Both keypoint and MANO-angle normalization files are exact merges of the per-sequence float64 accumulators. Global frame indices are generated only after all 627 sequences are CLEANED.

Conversion used 337 serial and 290 parallel-orchestrated physical sequences. The controller recorded 293 atomic claim events with 0 invalid overlapping claim lifecycles. Its exact source and audit report are preserved under Audit/orchestration/.

Build/final_manifest.json hashes every release-level artifact, and Build/release_ledger.sqlite3 records the exact source files, sequence manifests, camera episodes, anchor masks, and partial statistics. The fresh independent audit is published at Audit/independent_release_verification.json.

OakInk2 is human hand-object motion rather than robot control. Direct robot policy training still requires an explicit embodiment/retargeting decision. Source data and this derivative are CC BY-SA 4.0; MANO assets are not included and retain their separate license.