The dataset of the paper "Topology-Aware Optimization of Gaussian Primitives for Human-Centric Volumetric Videos".
Yuheng Jiang, Chengcheng Guo, Yize Wu, Yu Hong, Shengkun Zhu, Zhehao Shen, Yingliang Zhang, Shaohui Jiao, Zhuo Su, Lan Xu, Marc Habermann, Christian Theobalt.
SIGGRAPH Asia 2025
[Project Page] [Paper Link] [Dataset Link]
We have currently uploaded 7 sample cases of data with different people and diverse motions.
To extract videos from the dataset into a standard data format, first clone BackgroundMattingV2 and download the pre-trained model pytorch_resnet50.pth. Then, put pytorch_resnet50.pth to BackgroundMattingV2/ckpt.
git clone https://github.com/PeterL1n/BackgroundMattingV2
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
Use process.sh to extract frames, perform matting, and apply undistortion to the data:
bash process.sh
The overall file structure after processing is as follows:
├── image_white
│ ├── %d - The frame number, starts from 0.
│ │ └──%d.png - Masked RGB images for each view. view number starts from 0.
│ └── transforms.json - Camera extrinsics and intrinsics in NGP format.
│
├── image_white_undistortion
│ ├── %d - The frame number, starts from 0.
│ │ └──%d.png - Undistorted maksed RGB images for each view. view number starts from 0.
│ └── colmap/sparse/0 - Camera extrinsics and intrinsics in Gaussian Splatting format.
To run algorithms on single-frame data, you need to move the calibration files. For image_white under instant-ngp format, rename the frame folder to images and move transforms.json to the same directory as images. For image_white_undistortion in colmap format, rename the frame folder to images and move the sparse/0 in image_white_undistortion/colmap folder to the same directory as images.
Note on 3D Points: Since we used separately captured calibration data for camera calibration, the points3D.bin file in sparse/0 does not contain the point cloud of the actual scene. Therefore, we provide the sparse point cloud from the first frame of each sequence (reconstructed by 3D Gaussian Splatting) in the first_frame_pcds folder. A compressed archive of these point clouds is also available in the dataset download link.
The dataset is publicly accessible at Link. It is authorized for non-commercial use only. For commercial use, please contact Yuheng Jiang (nowheretrix123@gmail.com) and cc Chengcheng Guo(guochch2024@shanghaitech.edu.cn) to obtain necessary permissions.
If you find this work useful for your research, please cite us using the following BibTeX entry.
@misc{jiang2025topology,
title={Topology-Aware Optimization of Gaussian Primitives for Human-Centric Volumetric Videos},
author={Yuheng Jiang and Chengcheng Guo and Yize Wu and Yu Hong and Shengkun Zhu and Zhehao Shen and Yingliang Zhang and Shaohui Jiao and Zhuo Su and Lan Xu and Marc Habermann and Christian Theobalt},
year={2025},
eprint={2509.07653},
archivePrefix={arXiv},
primaryClass={cs.GR},
url={https://arxiv.org/abs/2509.07653},
}