# UniSH: Unifying Scene and Human Reconstruction in a Feed-Forward Pass **Mengfei Li**¹, **Peng Li**¹, **Zheng Zhang**², **Jiahao Lu**¹, **Chengfeng Zhao**¹, **Wei Xue**¹,
**Qifeng Liu**¹, **Sida Peng**³, **Wenxiao Zhang**¹, **Wenhan Luo**¹, **Yuan Liu**^1†, **Yike Guo**^1† ¹HKUST, ²BUPT, ³ZJU

--- ### TL;DR Given a monocular video as input, our UniSH is capable of jointly reconstructing scene and human in a single forward pass, enabling effective estimation of scene geometry, camera parameters and SMPL parameters.

## 🛠️ Installation We provide a **sudo-free** installation method that works on most Linux servers (including headless ones). ### Step 1: Clone Repository ```bash git clone https://github.com/murphylmf/UniSH.git cd UniSH ``` ### Step 2: Create Conda Environment This installs Python, system compilers, and OpenGL drivers. ```bash conda env create -f environment.yml conda activate unish ``` ### Step 3: Compile Dependencies This script compiles PyTorch3D, MMCV, and SAM2 from source using the compilers installed in Step 2. The environment has been tested on **CUDA 12.1** and **CUDA 11.8**. You can specify the CUDA version by passing it as an argument to the installation script. ```bash # Default (Auto-detect or 12.1) bash install.sh # For CUDA 11.8 bash install.sh 11.8 # For CUDA 12.1 bash install.sh 12.1 ``` ### Step 4: Download SMPL Models Please download the [SMPL](https://smpl.is.tue.mpg.de/) models and place them in the `body_models` folder. The directory structure should be organized as follows: ```text UniSH/ ├── body_models/ │ └── smpl/ │ └── smpl/ │ ├── SMPL_FEMALE.pkl │ ├── SMPL_MALE.pkl │ └── SMPL_NEUTRAL.pkl ``` ## 🚀 Quick Start (Inference) ### Run Inference Run the following command to reconstruct the scene and human from the video: ```bash python inference.py --output_dir inference_results/example --video_path examples/example_video.mp4 ``` Please refer to `inference.py` for more information about additional parameters. ## 📝 Citation If you find this code useful for your research, please consider citing our paper: ```bibtex @misc{li2026unishunifyingscenehuman, title={UniSH: Unifying Scene and Human Reconstruction in a Feed-Forward Pass}, author={Mengfei Li and Peng Li and Zheng Zhang and Jiahao Lu and Chengfeng Zhao and Wei Xue and Qifeng Liu and Sida Peng and Wenxiao Zhang and Wenhan Luo and Yuan Liu and Yike Guo}, year={2026}, eprint={2601.01222}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2601.01222}, } ``` ## 🙏 Acknowledgements We acknowledge the excellent contributions from the following projects: * [GVHMR](https://github.com/zju3dv/GVHMR) * [BEDLAM](https://bedlam.is.tue.mpg.de/) * [SMPL](https://smpl.is.tue.mpg.de/) * [VGGT](https://github.com/facebookresearch/vggt) * [Pi3](https://github.com/yyfz/Pi3) * [MoGe2](https://github.com/microsoft/moge) ## 📄 License This project is licensed under the Apache 2.0 License. See [LICENSE](LICENSE) for details.