RoGSplat: Learning Robust Generalizable Human Gaussian Splatting from Sparse Multi-View Images

Paper

This is official code of CVPR2025 paper RoGSplat: Learning Robust Generalizable Human Gaussian Splatting from Sparse Multi-View Images.

Installation

Set up python environment

sh ./install.sh

Set up dataset

RenderPeople

Please follow sherf to download the RenderPeople dataset. Unzip and rename it into renderpeople, split dataset into train and val.

Preprocessed SMPL

Download our estimated and preprocessed data from here and put it under renderpeople folder.

The final "renderpeople" folder should be like this:

renderpeople
  ├-- train
  ├-- val
  ├-- easymocap_smpl

Make a soft link:

ln -s /path/to/renderpeople ./renderpeople

Trainning

Train first stage:

python train_stage1.py

Then change stage1_ckpt to ./experiments/rp_xxxx/ckpt/rp_final.pth, change depth_ckpt to ./experiments/rp_xxxx/ckpt/rp_depth_latest.pth.

Then train second stage:

python train_stage2.py

Evaluation

python test_stage2.py --ckpt_path experiments/rp_xxxx

Citation

If you find this code useful for your research, please cite this:

@inproceedings{RoGSplat2025CVPR,
    title={{RoGSplat}: Learning Robust Generalizable Human Gaussian Splatting from Sparse Multi-View Images},
    author={Xiao, Junjin and Zhang, Qing and Nie, Yongwei and Zhu, Lei and Zheng, Wei-Shi},
    booktitle={CVPR},
    year={2025}
}