D3-rgb640: Ego-WMM RGB640

Public research snapshot (explicitly authorized 2026-11-03). D3 train 12,160 / validation 727. Existing historical three training exclusions retained. No resplitting.

All RGB videos are 640×640, 149 frames, 30fps. Original MP4 and normalized motion NPZ bytes are packed without recompression, regeneration, or new evaluations. Original split, captions, calibration provenance and fixed-generation GCS refs/SHA256 are retained in JSON metadata. The source data terms remain applicable; this upload does not grant a new license.

Layout

WebDataset TAR shards contain adjacent <source_sample_id>.mp4, .npz, .json. manifests/*.jsonl maps each row to its shard and members; SHARDS.json gives SHA256 and counts. The old original_row may reference960 sources: use the archived MP4s or top-level640 video_ref, never that old media path. Motion is the existing neutral-pivot camera9+SMPL147 raw156 representation.

from datasets import load_dataset
ds = load_dataset("X-Zhang/D3-rgb640", token=True, streaming=True)

For raw extraction, download a TAR shard with hf_hub_download(repo_type="dataset", ...), then extract the MP4/NPZ/JSON members. No model weights are included.

Dependencies include original action normalization statistics and the existing per-recording D3 Fisheye624 calibration asset/lookup. No new calibration was fitted.