chen-si-02/AnyHand-Dataset

🤲 AnyHand

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

NeurIPS 2026 · Evaluations and Datasets Track

6,337,536 released RGB-D samples · 47,438 hand shapes · 10,240 hand textures · 500K+ objects

Explore the project · Read the paper · Run the Colab demo

Official dataset release. Both train and test splits are available in this Hugging Face repository. The current release contains 6,189 standard TAR shards / 6,337,536 RGB-D samples, including 32 test shards / 32,768 test samples. The approximately 6.6M-image figure reported in the paper describes the research-scale dataset; the exact available inventory is listed below. Counts were verified against the published train/test manifests and HF file size/SHA-256 metadata on October 14, 2026.

Access: This public dataset uses automatic access approval. Sign in to Hugging Face, accept the dataset conditions, and authenticate before downloading. The original TAR format is retained and the built-in dataset viewer is disabled.

Overview

AnyHand is a large-scale synthetic dataset for 3D hand pose and mesh reconstruction, supporting both RGB-only and RGB-D learning. It brings together diverse hand articulations, realistic skin and forearm appearance, varied cameras and illumination, hand-object occlusion, and paired geometric supervision at a scale difficult to obtain from real capture.

The paper describes two complementary branches. The following table gives the scale reported in the paper; use the published train/test inventory below for the exact available samples:

Branch	Description	Images	Rendered scenes
AnyHand-Single	Isolated hands with attached forearms, broad pose/shape/appearance and viewpoint diversity	2.5M	1.25M
AnyHand-Interact	Hands manipulating objects, including natural grasps and object-induced occlusion	4.1M	2.05M
Total	Aligned RGB-D and rich geometric annotations	6.6M	3.30M

Each scene is rendered from two camera viewpoints. The dataset is designed to complement real-world training datasets, not necessarily to replace them: the paper shows that combining synthetic and real data improves hand reconstruction and generalization.

Published train/test inventory

Split	Repository directory	TAR shards	Samples
Train	data/train/single/	2,440	2,498,560
Train	data/train/interact-env/	1,630	1,669,120
Train	data/train/interact-indoor/	2,087	2,137,088
Train subtotal		6,157	6,304,768
Test	data/test/single-env/	8	8,192
Test	data/test/single-indoor/	8	8,192
Test	data/test/interact-env/	8	8,192
Test	data/test/interact-indoor/	8	8,192
Test subtotal		32	32,768
Total		6,189	6,337,536

Each standard TAR contains 1,024 paired samples. All four test categories have eight TARs each, named 000000.tar through 000007.tar. The test split contains 16,384 single-hand and 16,384 hand-object interaction samples across environment and indoor scenes.

The synthetic test split supports evaluation and debugging within this distribution. Keep it separate from training and report its results separately from the real-image benchmark protocols in the paper, including FreiHAND, HO-3D, and HO-Cap.

Use the training manifest and test manifest as the source of truth for file selection and reproducible experiments.

At a glance

Diversity / supervision	Scale or detail
Hand shapes	47,438 MANO shape configurations
Hand textures	10,240 appearances generated using Handy
Forearm textures	254 textures from SMPLitex
Background sources	1,270 indoor images and HDR environment maps
Hand-object interactions	500K+ objects represented in GraspXL / Objaverse
Ground truth and metadata	RGB, depth, foreground masks, bounding boxes, camera parameters, and 3D pose/shape annotations

Visual examples

AnyHand-Single — diverse hands, textures, scenes, and viewpoints

AnyHand-Interact — varied objects, grasps, and occlusions

More examples, the generation pipeline, and reconstruction comparisons are available on the project website.

Get the data

The dataset is distributed directly from this Hugging Face repository in its original TAR-shard format. You do not need a reconstruction network or MANO assets just to download or inspect RGB/depth images.

1. Authenticate and download the manifests

Request access on the dataset page, accept the conditions, and authenticate on your download machine:

python -m pip install -U huggingface_hub
hf auth login

Then download both inventories:

hf download chen-si-02/AnyHand-Dataset \
  manifests/shards.jsonl manifests/test_shards.jsonl \
  --repo-type dataset \
  --local-dir ./AnyHand

The HF client uses your locally saved token for authenticated downloads.

2. Choose the data you need

Entire test split — 32 TARs, about 4.16 GB:

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

One test shard:

hf download chen-si-02/AnyHand-Dataset \
  data/test/single-indoor/000000.tar \
  --repo-type dataset \
  --local-dir ./AnyHand

One training shard:

hf download chen-si-02/AnyHand-Dataset \
  data/train/single/000000.tar \
  --repo-type dataset \
  --local-dir ./AnyHand

Downloaded files preserve the repository-relative paths. For example, the test shard is saved at ./AnyHand/data/test/single-indoor/000000.tar. For training subsets, select exact paths from the manifest. The full train/test TAR inventory is approximately 768.3 GB.

3. Programmatic access

Read an RGB/depth pair without deserializing the annotation payload:

from io import BytesIO
import tarfile
from pathlib import Path

from huggingface_hub import hf_hub_download
from PIL import Image

path = hf_hub_download(
    repo_id="chen-si-02/AnyHand-Dataset",
    repo_type="dataset",
    filename="data/test/single-indoor/000000.tar",
)

with tarfile.open(path, "r") as tar:
    rgb_member = next(
        m for m in tar.getmembers()
        if m.isfile()
        and m.name.lower().endswith((".jpg", ".jpeg", ".png"))
        and not m.name.endswith((".depth.png", ".mask.png"))
    )
    stem = rgb_member.name.rsplit(".", 1)[0]

    with tar.extractfile(rgb_member) as fp:
        rgb = Image.open(BytesIO(fp.read())).convert("RGB")
    with tar.extractfile(tar.getmember(stem + ".depth.png")) as fp:
        raw_depth = Image.open(BytesIO(fp.read())).copy()

print("Sample:", Path(stem).name)
print("RGB size:", rgb.size, "Raw depth mode:", raw_depth.mode)
rgb.show()

Install Pillow with python -m pip install pillow. The depth image contains raw integer values; convert them using the annotation's per-sample depth scale and invalid-depth code described below. Read legacy labels with a reviewed, restricted/validated loader.

Data organization and annotation format

Repository layout

Both splits are organized under data/, with the test folder parallel to the training folder:

AnyHand-Dataset/
├── README.md
├── manifests/
│   ├── shards.jsonl                # training inventory
│   ├── test_shards.jsonl           # test inventory
│   └── missing_tar_ids.txt
├── transfer_summary.json
├── transfer_summary_test.json
├── release_audit.json
└── data/
    ├── train/
    │   ├── single/
    │   ├── interact-env/
    │   └── interact-indoor/
    └── test/
        ├── single-env/
        ├── single-indoor/
        ├── interact-env/
        └── interact-indoor/

Training source labels map as follows: dpx → single, graspxl_env → interact-env, and graspxl_indoor → interact-indoor. Test source labels map as dpx_envmap → single-env, dpx_indoor → single-indoor, grasp_envmap / graspxl_envmap → interact-env, and grasp_indoor / graspxl_indoor → interact-indoor. These are storage categories; the scientific branches remain AnyHand-Single and AnyHand-Interact.

Each published standard TAR shard currently contains 1,024 paired samples. Files belonging to one sample share a common basename:

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

The dataset contains geometric annotations such as 2D/3D hand keypoints, camera information, and hand pose/shape parameters. For the fields explicitly documented in the released TAR reader format:

Item	Convention
keypoints_2d	Full-image pixel coordinates with confidence values
keypoints_3d	Camera-space meters with confidence values
Camera axes	+X right, +Y down, +Z forward
cam_t	Translation is already reflected in stored 3D joints; do not add it again
depth_scale_to_m	Multiply raw uint16 depth values by this per-sample factor to obtain meters
depth_invalid_code	Identifies depth pixels to treat as invalid
Confidence	Not a guaranteed or independently verified visibility/occlusion mask

Depth caveat. Foreground hand/arm depth is produced by rendering. Background depth is estimated using MoGe-2 and composited in camera space, so the dense depth image is not uniformly exact ground truth, particularly near foreground/background boundaries. Foreground masks can be used to focus losses on relevant pixels when available.

Security note. The .data.pyd files use a legacy NumPy-pickle representation rather than JSON. Only load such annotations from a trusted source with an appropriately restricted/validated loader; arbitrary pickle loading can execute code. The ordinary image/depth files are accessible without deserializing the labels.

Integrity and release tracking

The publication pipeline validates RGB, depth, and annotation payloads, checks image readability, TAR termination, and optional masks, and records sample counts and provenance. Each published TAR's byte size and SHA-256 were confirmed against HF metadata. Files that fail validation are excluded rather than published as partial shards.

The training manifest and test manifest list paths, sizes, sample counts, SHA-256, and source identifiers. transfer_summary.json and transfer_summary_test.json record the completed split-specific transfers.

Training numbering: train/single has consecutive IDs 000000–002439. The indoor-interaction numeric sequence is consecutive 000000–002086, with filenames switching from 001089.tar to indoor_001090.tar. The environment-interaction series spans 000000–001919 with 290 absent IDs: five every 25 from 000470–000474 through 001895–001899. These gaps are also absent from the pinned source catalogue. Use manifest paths rather than generating filenames by incrementing numbers. See the training release audit and complete gap list.

Source provenance is retained from the pinned Harvard Dataverse catalogues: training, DOI 10.7910/DVN/VBYEK3, and test, DOI 10.7910/DVN/B5HHNR, both recorded as version :draft. The October 14 card update compared current HF file size/SHA-256 metadata with both manifests; it did not re-download and decode every TAR or re-audit the live Dataverse catalogues.

For reproducible experiments, pin an HF repository revision and retain the relevant manifests and release metadata with your experiment configuration.

Dataset creation

Our generation pipeline combines the following components:

1. Pose and shape sampling. MANO shapes are sampled from empirical distributions of real hand datasets; articulated poses come from the DPoser-Hand diffusion prior.
2. Appearance. Handy supplies rich hand textures, with additional appearance variation; SMPL-H forearms with SMPLitex textures provide realistic wrist/arm continuity.
3. Hand-object scenes. GraspXL simulation provides contact-consistent hand-object motions, paired with textured assets from Objaverse.
4. Rendering and cameras. SAPIEN renders hands and scenes with randomized viewpoints, intrinsics, backgrounds, and lighting; each scene yields two views.
5. RGB-D supervision. Rendered foreground depth and MoGe-2 background estimates are composed to produce aligned RGB and dense depth, along with geometric labels.

For the implementation and detailed distributions, please see Section 3 and Appendix C of the paper and the project page.

Results reported in the paper

In controlled experiments, AnyHand is added to the training data of existing methods while preserving the baseline architectures and training protocols. Representative PA-MPJPE results (millimeters; lower is better) are:

Benchmark	Model	Original training	Training with AnyHand
FreiHAND	HaMeR	6.000	5.545
HO-3D v2	WiLoR	7.500	7.355
HO-Cap (zero-shot)	HaMeR	4.940	4.660

The paper also studies RGB-D depth fusion, synthetic-data scaling, and real/synthetic mixing. These numbers are research results for the configurations reported in the paper, not precomputed metrics attached to every published TAR shard. See the paper for full protocols and additional benchmarks.

Intended use and limitations

Intended uses: research into 3D hand pose/mesh reconstruction, hand-object interactions, RGB-D supervision, synthetic-to-real generalization, and dataset-scaling or data-mixture strategies. The dataset also supports controlled analysis of appearance, pose, and viewpoint diversity.

Limitations and considerations:

• Synthetic-to-real gap: the coverage of rendered appearances, hands, objects, lighting, and behaviors does not fully represent real-world imagery.
• Pose realism: the DPoser-Hand prior and biomechanical filtering improve plausibility but may still admit anatomically unlikely configurations.
• Depth quality: the background portion of a dense depth map is estimated and composited, not measured by a real depth sensor; boundary artifacts are possible.
• Training recipe matters: the paper finds that appropriately mixing AnyHand with real data works better than using synthetic samples alone in the studied setup.
• Test-split scope: synthetic test results are separate from real-image benchmark results. No disjointness guarantee for exact subject, texture, or object identities across splits is asserted by this card.
• Asset and model conditions: constituent resources (including MANO-related assets, textures, and Objaverse objects) may carry their own terms. Users are responsible for complying with applicable third-party rights and restrictions.

See Appendix F for the limitations described in the publication.

License and citation

Dataset license: Creative Commons Attribution–NonCommercial–NoDerivatives 4.0 International (CC BY-NC-ND 4.0). This license applies to the dataset release, and should not be confused with the licenses of code, pretrained models, or third-party materials.

If you use AnyHand in your work, please cite the NeurIPS 2026 paper:

@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}
}

Links: Project website · Paper · Code · Checkpoints · Interactive demo · Dataset discussions

For questions or release issues, please use the GitHub issue tracker or the Hugging Face dataset discussions.