---
license: cc-by-nc-4.0
task_categories:
- image-to-3d
tags:
- hand-object-reconstruction
- hand-pose-estimation
- 3d-reconstruction
- multi-view
- synthetic
- hand
- MANO
pretty_name: HGGT Synthetic Dataset
configs:
- config_name: preview
  data_files:
  - split: preview
    path: preview/**
---

# HGGT Synthetic Dataset

This is the synthetic multi-view hand-object interaction dataset introduced in:

> **HGGT: Robust and Flexible 3D Hand Mesh Reconstruction from Uncalibrated Images**  
> Yumeng Liu, Xiao-Xiao Long, Marc Habermann, Xuanze Yang, Cheng Lin, Yuan Liu, Yuexin Ma, Wenping Wang, Ligang Liu  
> [[Paper]](https://arxiv.org/abs/2603.23997) · [[Project Page]](https://lym29.github.io/HGGT/) · [[Code]](https://github.com/lym29/HGGT)

The dataset contains diverse photorealistic hand-object interactions rendered with randomized camera viewpoints, providing critical viewpoint diversity absent in real-world captures. It is used together with real monocular and real multi-view data under a mixed-data training strategy.

## Dataset Structure

```
catmint123/HGGT-synthetic-data/
├── small/
│   ├── shard_00000.tar   # ~2 GB each, contains ~50 object uuid dirs
│   ├── shard_00001.tar
│   └── ...
├── medium/
│   ├── shard_00000.tar
│   └── ...
├── large/
│   ├── shard_00000.tar
│   └── ...
├── splits/
│   ├── train.txt         # object uuids for training
│   └── test.txt          # object uuids for testing
└── invalid_masks_train.json
```

`small`, `medium`, and `large` follow the object scale categorization from GraspXL, where object meshes are grouped by their physical size. Each tar shard contains a set of object uuid directories with the following structure:

```
<object_uuid>/
└── <motion_id>/            # GraspXL motion sequence index (e.g. 1, 2, ...)
    └── sequence_<XXXXXX>/  # a sampled frame from the motion
        ├── images/         # multi-view rendered RGB images (000000.png, 000001.png, ...)
        ├── masks/          # hand and object segmentation masks per view
        ├── multi_depth_gt/ # ground-truth depth maps per view (.npz)
        ├── camera/
        │   ├── extrinsics/ # per-view camera extrinsic matrices (.txt)
        │   └── intrinsics.txt
        └── info.json       # metadata (motion path, frame id, MANO params, object paths)
```

## Download

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="catmint123/HGGT-synthetic-data",
    repo_type="dataset",
    local_dir="data/hggt_synthetic",
)
```

To download a specific split only (e.g. `small`):

```python
snapshot_download(
    repo_id="catmint123/HGGT-synthetic-data",
    repo_type="dataset",
    local_dir="data/hggt_synthetic",
    allow_patterns="small/*",
)
```

After downloading, extract the tar shards:

```bash
cd data/hggt_synthetic/small
for f in *.tar; do tar -xf "$f"; done
```

## Usage

The `splits/train.txt` and `splits/test.txt` files list the object uuid directories used for training and testing respectively. 

Please refer to the [HGGT code repository](https://github.com/lym29/HGGT) for data loading and training scripts.

## Data Sources

This dataset is built upon the following works:

- **3D object assets**: Object meshes are sourced from [Objaverse](https://objaverse.allenai.org/) (Deitke et al., CVPR 2023), a large-scale dataset of 800K+ annotated 3D objects.
- **Hand-object interaction data**: Hand-object interaction configurations and grasping motions are sourced from [GraspXL](https://github.com/zdchan/GraspXL) (Zhang et al., ECCV 2024), which provides large-scale grasping motion data for 500k+ objects with diverse dexterous hands.
- **Hand texture materials**: Hand appearance textures are sourced from [DART](https://github.com/DART2022/DART) (Gao et al., NeurIPS 2022), an articulated hand model with diverse accessories and rich textures.

Please make sure to also cite these works if you use this dataset.

## License

This dataset is licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). You are free to share and adapt the data for non-commercial purposes, provided appropriate credit is given.

This dataset builds upon [Objaverse](https://objaverse.allenai.org/) (CC BY 4.0), [GraspXL](https://github.com/zdchan/GraspXL) (CC BY-NC 4.0), and [DART](https://github.com/DART2022/DART). Please review and comply with their respective licenses when using this dataset.

## Citation

```bibtex
@article{liu2026hggt,
  title={HGGT: Robust and Flexible 3D Hand Mesh Reconstruction from Uncalibrated Images},
  author={Liu, Yumeng and Long, Xiao-Xiao and Habermann, Marc and Yang, Xuanze and Lin, Cheng and Liu, Yuan and Ma, Yuexin and Wang, Wenping and Liu, Ligang},
  journal={arXiv preprint arXiv:2603.23997},
  year={2026}
}

@inproceedings{deitke2023objaverse,
  title={Objaverse: A universe of annotated 3d objects},
  author={Deitke, Matt and Schwenk, Dustin and Salvador, Jordi and Weihs, Luca and Michel, Oscar and VanderBilt, Eli and Schmidt, Ludwig and Ehsani, Kiana and Kembhavi, Aniruddha and Farhadi, Ali},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  pages={13142--13153},
  year={2023}
}

@inProceedings{zhang2024graspxl,
  title={{GraspXL}: Generating Grasping Motions for Diverse Objects at Scale},
  author={Zhang, Hui and Christen, Sammy and Fan, Zicong and Hilliges, Otmar and Song, Jie},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2024}
}

@inproceedings{gao2022dart,
    title={{DART: Articulated Hand Model with Diverse Accessories and Rich Textures}},
    author={Daiheng Gao and Yuliang Xiu and Kailin Li and Lixin Yang and Feng Wang and Peng Zhang and Bang Zhang and Cewu Lu and Ping Tan},
    booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
    year={2022},
}
```