SizeGS: Size-aware Compression of 3D Gaussian Splatting via Mixed Integer Programming

1. Cloning the Repository

git clone https://github.com/mmlab-sigs/sizegs --recursive

2. Install

2.1 Hardware and Software Requirements

• CUDA-ready GPU with Compute Capability 7.0+
• Ubuntu >= 18.04
• Conda (recommended for easy setup)
• C++ Compiler for PyTorch extensions (we recommend Visual Studio 2019 for Windows)
• CUDA >= 11.6
• C++ Compiler and CUDA SDK must be compatible

2.2 Setup

Our provided install method is based on Conda package and environment management:

CUDA 12.1/12.4:

conda create -n sizegs python=3.10
conda activate sizegs
pip install torch==2.2.2+cu121 torchvision==0.17.2+cu121 torchaudio==2.2.2+cu121 --index-url https://download.pytorch.org/whl/cu121
pip install torch-scatter torch-cluster -f https://data.pyg.org/whl/torch-2.2.2+cu121.html
pip install tqdm plyfile einops wandb lpips laspy colorama jaxtyping opencv-python tensorboard loguru pulp Ninja open3d torchac
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn
pip install submodules/octree
pip install submodules/quant
pip install submodules/seg-quant
pip install numpy==1.26.4

2.3 Preparing dataset and pre-trained ScaffoldGSes

2.3.1 Dataset

First, create a data/ folder inside the project path by

mkdir data

The data structure will be organised as follows:

data/
├── dataset_name
│   ├── scene1/
│   │   ├── images
│   │   │   ├── IMG_0.jpg
│   │   │   ├── IMG_1.jpg
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
│   ├── scene2/
│   │   ├── images
│   │   │   ├── IMG_0.jpg
│   │   │   ├── IMG_1.jpg
│   │   │   ├── ...
│   │   ├── sparse/
│   │       └──0/
...

2.3.2 Pretrained ScaffoldGSes

Then, create a outputs/ folder inside the project path by

mkdir outputs

Take T&T as an example, the outputs/tandt structure should be organized as follows. An example lego ScaffoldGS can be downloaded from here. These pretrained ScaffoldGSes serve at the input of our code.

.
├── train
│   └── baseline
│       └── scaffold
│           ├── backup
│           ├── cameras.json
│           ├── cfg_args
│           ├── input.ply
│           ├── outputs.log
│           ├── per_view.json
│           ├── point_cloud
│           ├── results.json
│           └── test
└── truck
    └── baseline
        └── scaffold
            ├── backup
            ├── cameras.json
            ├── cfg_args
            ├── input.ply
            ├── outputs.log
            ├── per_view.json
            ├── point_cloud
            ├── results.json
            └── test

At last change the path/to/outputs/ and path/to/data/ in meson.sh with your path

3. Running

Use torchac

• Check size_ac_mip_360.sh.

Use LZ77 (zip)

• Check size_mip_360.sh.

• Check size_tandt.sh.

• Check size_db.sh.

TODO

• lint the code.

4. Citation

If you find our work helpful, please consider citing:

@inproceedings{xie2024sizegs,
    title={SizeGS: Size-aware Compression of 3D Gaussian Splatting via Mixed Integer Programming},
    author={Xie, Shuzhao and Liu, Jiahang and Zhang, Weixiang and Ge, Shijia and Pan, Sicheng and Tang, Chen and Bai, Yunpeng and Zhang, Cong and Fan, Xiaoyi and Wang, Zhi},
    booktitle={ACM MM},
    year={2025}
}

5. Acknowledgments

We sincerely thank the anonymous reviewers from ICLR, CVPR, and ACM MM for their valuable feedback and suggestions. We also thank our lab mates for their help in improving the manuscript. This work is supported in part by National Key Research and Development Project of China (Grant No. 2023YFF0905502), National Natural Science Foundation of China (Grant No. 92467204 and 62472249), and Shenzhen Science and Technology Program (Grant No. JCYJ20220818101014030 and KJZD20240903102300001).