MANUS-HaGRID: HaGRID-derived Multimodal Annotated Naturalistic Hand Understanding Dataset

MANUS-HaGRID is the HaGRID/HaGRIDv2-derived subset of the Multimodal Annotated Naturalistic Hand Understanding (MANUS) dataset family. It provides multimodal annotations for naturalistic hand gesture understanding, including RGB images, hand crops, depth maps, 2D bounding boxes, estimated MANO-style hand mesh metadata, and multi-view mesh renderings where available.

This repository contains only the HaGRID-derived MANUS subsets: hagrid_v1 and hagrid_v2. DexYCB-derived data is released separately as MANUS-DexYCB.

Dataset Features in This Release

MANUS-HaGRID provides the following modalities where available:

• full-scene RGB image;
• full-scene depth map at 512 x 512 resolution;
• hand-region RGB crop at 225 x 225 resolution, when available;
• hand-region depth map at 225 x 225 resolution;
• 2D bounding boxes for hand regions;
• estimated MANO-style hand mesh annotations or related mesh metadata;
• rendered mesh views from canonical viewpoints including front, rear, left, right, top, and bottom, when available;
• companion text prompts describing image content, provided as split-level prompt files listed below;
• per-sample metadata recording source dataset, source id, split, gesture label, available modalities, and license metadata in the public manifest.

Text prompts are provided as companion split-level files at the repository root. They are not embedded inside the existing WebDataset tar shards; use the prompt files listed below when text conditioning is needed.

Configs and Splits

Config/Split	Samples	WebDataset shards	Size
hagrid_v1/train	28,814	18	16.73 GB
hagrid_v1/test	7,623	8	7.20 GB
hagrid_v2/train	36,123	33	30.27 GB
hagrid_v2/test	7,916	7	6.34 GB

Total canonical samples in this release: 80,476. Total WebDataset tar size: 60.55 GB.

MANUS-HaGRID Release Scope

Included:

• hagrid_v1/train: based on the HaGRID V1 training subset selected by the MANUS processing pipeline.
• hagrid_v1/test: based on the HaGRID V1 test subset selected by the MANUS processing pipeline.
• hagrid_v2/train: based on the HaGRIDv2 training subset selected by the MANUS processing pipeline.
• hagrid_v2/test: based on the HaGRIDv2 test subset selected by the MANUS processing pipeline.

Not included:

• MANUS-DexYCB, which is released in a separate repository.

Data Format

Data is stored as uncompressed WebDataset tar shards. Files belonging to the same sample share the same prefix:

v2_test_grabbing_00000001.json
v2_test_grabbing_00000001.full_rgb.png
v2_test_grabbing_00000001.full_depth.png
v2_test_grabbing_00000001.hand_rgb.png
v2_test_grabbing_00000001.hand_depth.png
v2_test_grabbing_00000001.bbox.json
v2_test_grabbing_00000001.normal.json
v2_test_grabbing_00000001.mesh_front.png

The per-sample .json member records source id, split, gesture label, and the list of files available for that sample. The public manifest adds source/license fields such as source_dataset, source_license_id, contains_source_pixels, and processing_pipeline_version. See docs/data_structure.md and docs/annotation_format.md for details.

Known Modality Differences in This Release

• V1 test has no usable hand RGB source package, so hand_rgb.png is absent for that split.
• V2 test has 7,916 canonical OG samples, but only 7,911 samples have hand crops, bbox JSON, normal JSON, and six-view mesh assets in the provided source packages.
• V1 train has 28,814 canonical samples; normal.json is present for 28,780 of them.
• V1 train six-view mesh source contains 28,860 images per view; this release filters to the 28,814 canonical OG ids.

Text Prompt Files

The following companion prompt files have been added at the repository root:

File	Split	Records/Lines	Format	Notes
HaGRID_V1_train_prompt.jsonl	hagrid_v1/train	28,814	JSONL	One JSON object per sample with conditioning_image, image, text, and label.
HaGRID_V1_test_prompt.txt	hagrid_v1/test	7,623	TXT	One text prompt per line, aligned to the canonical V1 test order.
HaGRID_V1_test_gesture_label.txt	hagrid_v1/test	7,623	TXT	One gesture label per line, aligned to the canonical V1 test order.
HaGRID_V2_train_prompt.jsonl	hagrid_v2/train	36,117	JSONL	One JSON object per available prompted sample with conditioning_image, image, text, and label.
HaGRID_V2_test_prompt.txt	hagrid_v2/test	7,909	TXT	One text prompt per line for available prompted V2 test samples.
HaGRID_V2_test_gesture_label.txt	hagrid_v2/test	7,909	TXT	One gesture label per line for available prompted V2 test samples.

Prompt and gesture-label files are optional companion annotations and are not packed into the WebDataset shards. The V2 prompt/label files contain fewer records than the canonical V2 sample counts (hagrid_v2/train: 36,123; hagrid_v2/test: 7,916), so users should treat these companion annotations as available-where-present rather than required fields for every V2 sample.

Metadata Files

• metadata/hagrid_v1_manifest.parquet
• metadata/hagrid_v2_manifest.parquet
• metadata/hagrid_all_manifest.parquet
• metadata/source_licenses.csv
• metadata/release_schema.json
• metadata/release_report.json
• metadata/source_package_audit.csv
• metadata/checksums.sha256
• HaGRID_V1_train_prompt.jsonl
• HaGRID_V1_test_prompt.txt
• HaGRID_V1_test_gesture_label.txt
• HaGRID_V2_train_prompt.jsonl
• HaGRID_V2_test_prompt.txt
• HaGRID_V2_test_gesture_label.txt

The public manifest files do not contain local absolute source paths. They contain portable source and audit metadata, including sample id, source id, source license, available modalities, and shard path.

Loading

See docs/loading_examples.md.

Source Dataset Reference

This release is derived from HaGRID/HaGRIDv2. Please acknowledge the upstream HaGRID datasets when using MANUS-HaGRID.

@misc{nuzhdin2024hagridv21mimagesstatic,
  title={HaGRIDv2: 1M Images for Static and Dynamic Hand Gesture Recognition},
  author={Anton Nuzhdin and Alexander Nagaev and Alexander Sautin and Alexander Kapitanov and Karina Kvanchiani},
  year={2024},
  eprint={2412.01508},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2412.01508}
}

@inproceedings{kapitanov2024hagrid,
  title={HaGRID--HAnd Gesture Recognition Image Dataset},
  author={Kapitanov, Alexander and Kvanchiani, Karina and Nagaev, Alexander and Kraynov, Roman and Makhliarchuk, Andrei},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={4572--4581},
  year={2024}
}

Citation

If you use MANUS-HaGRID, please cite both the MANUS / HanDrawer work and the upstream HaGRID / HaGRIDv2 datasets.

MANUS / HanDrawer

@misc{fu2025handrawer,
  title         = {HanDrawer: Leveraging Spatial Information to Render Realistic Hands Using a Conditional Diffusion Model in Single Stage},
  author        = {Qifan Fu and Xu Chen and Muhammad Asad and Shanxin Yuan and Changjae Oh and Gregory Slabaugh},
  year          = {2025},
  eprint        = {2503.02127},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2503.02127},
  url           = {https://arxiv.org/abs/2503.02127}
}

HaGRID / HaGRIDv2

Use the HaGRID citations listed in the source dataset reference above.

License

MANUS-HaGRID contains source images and image derivatives from HaGRID/HaGRIDv2. The source RGB images, resized images, cropped hand images, and visual derivatives remain governed by the upstream HaGRID-specific license.

The upstream HaGRID repository states that HaGRID is licensed under a variant of the Creative Commons Attribution-ShareAlike 4.0 International License. Please see the upstream HaGRID license for the exact terms.

This repository uses license: other in the Hugging Face metadata because the upstream license is described as a HaGRID-specific CC BY-SA 4.0 variant rather than a plain standard license identifier.

This repository does not grant any rights beyond those granted by the upstream HaGRID/HaGRIDv2 license. Users are responsible for complying with the upstream license, attribution requirements, change notices, and any other applicable rights.

Intended Uses

MANUS-HaGRID is intended for research on hand gesture recognition, hand-aware image generation, 2D/3D hand understanding, depth estimation, and multimodal hand representation learning.

Out-of-Scope Uses

This dataset should not be used for biometric identification, person recognition, surveillance, or privacy-invasive applications. Users must comply with the upstream HaGRID license and applicable laws.

Privacy and Takedown

MANUS-HaGRID is derived from images of people showing hand gestures. Some images may include identifiable visual context. If you are a rights holder or data subject and request removal of a sample, please contact the dataset maintainers through the repository contact channel.

Related MANUS Repositories

• MANUS landing page: MANUS
• MANUS-DexYCB: DexYCB-derived MANUS release

This repository contains only HaGRID-derived data.