chen-si-02/AnyHand-Dataset

AnyHand

Original AnyHand standard TAR shards for synthetic hand-pose research. This is a distribution mirror of the Harvard Dataverse source, persistent ID doi:10.7910/DVN/VBYEK3, source version :draft.

Release status

Transfer in progress or partial. Uploaded samples: 4,253,696. Uploaded shards: 4,154. Quarantined shards: 0; quarantined outer bundles: 0. Final remote audit commit: not yet audited. Failed items are excluded, not replaced with partial shards. Check transfer_summary.json before treating this mirror as a complete release.

Files and integrity

Each published TAR contains exactly 1024 samples. Every sample's RGB, depth, and annotation were decoded and checked locally; PNG checksums and TAR end markers were checked. Optional masks were checked when present. Upload size and SHA-256 were confirmed through HF metadata.

Files preserve source split folders, for example data/train/single/000000.tar. Category folders are renamed: dpx becomes single, gxl-indoor becomes interact-indoor, and gxl-ourdoor/gxl-outdoor become interact-outdoor. manifests/shards.jsonl lists paths, sizes, sample counts, SHA-256, and source directory/bundle/member identifiers. Split labels are copied from explicit source folders when available; use the original release's split definitions.

Format

00000000.jpg           RGB image (some shards may use PNG/JPEG)
00000000.depth.png     uint16 depth
00000000.mask.png      optional grayscale mask
00000000.data.pyd      legacy NumPy pickle containing [annotation_dict]

keypoints_2d are full-image pixel coordinates with confidence; keypoints_3d are camera-space metres with confidence. The camera axes are X right, Y down, Z forward. Do not add cam_t again to stored 3D joints. Multiply raw depth by depth_scale_to_m to get metres; the invalid depth code is given by depth_invalid_code. Confidence is not a verified occlusion mask. Read original labels without executing unrestricted pickle code.

Access

Use AnyHand Toolkit, which includes an RGB/depth/keypoint notebook and a lightweight TAR reader:

python -m pip install git+https://github.com/chen-si-cs/anyhand-toolkit.git
python -m pip install huggingface_hub
hf download chen-si-02/AnyHand-Dataset --repo-type dataset --include "data/**/*.tar" --local-dir data

from anyhand_toolkit import Shard
with Shard("data/data/train/single/000000.tar") as shard:
    sample = shard[0]
rgb, depth_m, labels = sample.rgb, sample.depth_m, sample.annotation

Replace the example with a path from the manifest. The download command preserves the repository's data/ prefix. For a small subset, pass exact paths to hf download rather than downloading all TARs. Automatic dataset preview is disabled because the original annotation format uses legacy pickle.

License and citation

Dataset license: cc-by-nc-nd-4.0. This is separate from the toolkit's MIT code license. Follow the original release's licensing and third-party asset conditions. Source: doi:10.7910/DVN/VBYEK3, version :draft.

For citation, use the citation supplied by the original Dataverse release.

Uses and limitations

This distribution supports hand-pose research using rendered RGB, depth, and stored labels. Synthetic appearance and pose coverage may differ from real images. The source release remains authoritative for generation methods, sampling distributions, evaluation splits, and detailed scientific limitations.