/CompGS

source code for our work "CompGS: Efficient 3D Scene Representation via Compressed Gaussian Splatting"

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

source code for our work "CompGS: Efficient 3D Scene Representation via Compressed Gaussian Splatting"

rd_curve

Overview

Overview

Preparation

Download dataset

We evaluate our method on the following datasets:

Please download the datasets and unzip them.

Installation dependencies

Requirments: Python 3.10 and above

  1. Clone the repository

    git clone git@github.com:LiuXiangrui/CompGS.git

  2. Install Pytorch and other dependencies specified in the requirements.txt

    (Windows only) Please install CompressAI from source since wheels are available for Linux and MacOS.

  3. Install KNN and Splatting modules

    SET DISTUTILS_USE_SDK=1 # Windows only
cd CompGS/submodules/knn_dist
pip install .
cd ../diff-gaussian-rasterization
pip install .

  4. Compile GPCC

    Please download the GPCC source code from here and then compile it.

Train and evaluate our method

  • Please refer to Sripts/derive_train_eval_scripts for training and evaluation scripts.

  • Note that you need to modify the paths in the scripts to your own paths.

    • trainer_path: path of the Train.py file (absolute path is recommended)
    • gpcc_codec_path: path of the GPCC codec executable file (absolute path is recommended)
    • dataset_root: path of the dataset (absolute path is recommended)
    • experiments_root: path to save the experiments (absolute path is recommended)
    • config_path: path of the config file (absolute path is recommended)

  • Then you can run the script run.bat of run.sh (depends on your system) to train and evaluate our method.

  • After the evaluation, you can find the results in the experiments_root/config_name/experiment_folder/eval/results.json file. The results will include:

    • decompression_time: time for decompression in minutes.
    • per_view: quality scores for each view.
    • average: average quality scores.
    • render_time: time for rendering each view in milliseconds.
    • num_gaussians: number of anchor primitives.
    • training_time: time for training in minutes.
    • file_size: bitstream size in bytes, including size of each component.
    • compression_time: time for compression in minutes.

Render based on compressed models

You can render novel views from compressed models by running Test.py

Here is an example of rendering on the Train scene of Tanks&Temples dataset:

python Test.py --dataset_root D:\T2T\Datasets\Train --experiment_root D:\T2T\Results\Train\Lambda0_001\2024-04-07_16-57 \\
               --voxel_size=0.01 --derive_factor=10 --ref_feats_dim=32 --ref_hyper_dim=4 --res_feats_dim=8 --res_hyper_dim=1 \\
               --gpcc_codec_path=C:\Users\XiangruiLIU\Desktop\mpeg-pcc-tmc13\build\tmc3\Release\tmc3.exe

Acknowledgment

❤️❤️❤️ This work is implemented based on the following projects.

We really appreciate their wonderful open-source work!

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{liu2024compgs,
  title={CompGS: Efficient 3D Scene Representation via Compressed Gaussian Splatting},
  author={Liu, Xiangrui and Wu, Xinju and Zhang, Pingping and Wang, Shiqi and Li, Zhu and Kwong, Sam},
  booktitle={Proceedings of the 32nd ACM International Conference on Multimedia},
  year={2024}
}