chen-si-02/AnyHand-Dataset

🤲 AnyHand

A Large-Scale Synthetic Dataset for RGB(-D) Hand Pose Estimation

NeurIPS 2026 · Evaluations and Datasets Track

6.6M-scale RGB-D dataset · 47,438 hand shapes · 10,240 hand textures · 500K+ objects

AnyHand v1 provides paired RGB-D images and geometric annotations for 3D hand pose and mesh reconstruction. It combines diverse poses, shapes, appearance, lighting, and viewpoints across AnyHand-Single and AnyHand-Interact, supporting both RGB-only and RGB-D research.

Branch	Content	Scale
AnyHand-Single	Hands with forearms across diverse scenes and viewpoints	~2.5M images
AnyHand-Interact	Object manipulation, grasps, and occlusions	~4.1M images

Each scene is rendered from two camera viewpoints. See the paper and project for generation details, experiments, and benchmarks.

Visual examples

AnyHand-Single

AnyHand-Interact

Download

The train and test splits are public and distributed as standard TAR shards. Start with one test shard:

python -m pip install -U huggingface_hub
hf download chen-si-02/AnyHand-Dataset data/test/single-indoor/000000.tar --repo-type dataset --local-dir ./AnyHand

To download the complete test split:

hf download chen-si-02/AnyHand-Dataset --repo-type dataset --include 'data/test/*/*.tar' --local-dir ./AnyHand

Browse data/train and data/test, or select paths from the training manifest and test manifest. Each manifest records the exact file paths, sample counts, byte sizes, and SHA-256 checksums. Use these paths directly when selecting training subsets.

Split	Categories
Train	single, interact-env, interact-indoor
Test	single-env, single-indoor, interact-env, interact-indoor

Downloads preserve paths such as AnyHand/data/test/single-indoor/000000.tar. For reproducibility, add --revision <commit> to downloads and retain the corresponding manifests.

Read and visualize

Use the AnyHand Toolkit for reading, validation, checksummed downloads, and RGB-D/3D visualization. The core reader needs only NumPy and Pillow; model weights and MANO assets are unnecessary.

python -m pip install "git+https://github.com/chen-si-cs/anyhand-toolkit.git"

from anyhand_toolkit import Shard

with Shard("AnyHand/data/test/single-indoor/000000.tar") as shard:
    sample = shard[0]

rgb = sample.rgb
depth_m = sample.depth_m
labels = sample.annotation

For interactive examples and command-line previews, see the toolkit's notebook and documentation.

Format and conventions

Each standard shard contains 1,024 samples. A sample shares a basename across its files:

00000000.jpg          RGB image (.jpg, .jpeg, or .png)
00000000.depth.png    uint16 depth
00000000.mask.png     optional foreground mask
00000000.data.pyd     NumPy-pickle annotation dictionary/list

• 2D joints: full-image pixel coordinates, with confidence values.
• 3D joints: camera-space metres; axes are X right, Y down, Z forward. Stored joints already include cam_t.
• Depth: multiply raw values by depth_scale_to_m; treat depth_invalid_code as invalid. The toolkit handles both.
• Labels: include pose/shape and camera metadata. Confidence values are not a verified visibility mask.

See the toolkit's format reference for field details. Legacy .data.pyd annotations use pickle: load trusted release files with the toolkit's restricted reader rather than unrestricted pickle loading.

Use and limitations

AnyHand supports research into hand reconstruction, hand-object interactions, RGB-D learning, and synthetic-to-real generalization. Mixing synthetic and real data improves performance in the paper's studied settings.

• Rendered appearance and poses retain a synthetic-to-real gap and may include implausible configurations.
• Foreground depth is rendered; background depth is estimated with MoGe-2 and composited, so dense depth is not uniformly exact ground truth. Boundary artifacts are possible.
• Synthetic test results should be reported separately from real-image benchmarks. This card does not assert disjoint hand, texture, or object identities across splits.
• Third-party assets, including MANO-related materials, textures, and objects, may have separate terms.

See the paper for protocols and limitations. For questions, use dataset discussions.

License and citation

Dataset: CC BY-NC-ND 4.0. Toolkit code is separately licensed under MIT. Model and third-party asset terms are separate from the dataset license.

@inproceedings{si2026anyhand,
  title     = {{AnyHand}: A Large-Scale Synthetic Dataset for {RGB(-D)} Hand Pose Estimation},
  author    = {Si, Chen and Liu, Yulin and Ai, Bo and Xie, Jianwen and
               Potamias, Rolandos Alexandros and Zheng, Chuanxia and Su, Hao},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}