HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting

This repository contains the code for the paper titled "HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting".

[paper] [project page] [code] [checkpoints] Currently, only the inference part of the code is open sourced to demonstrate some implementation details and effects. The training part of the code is not open sourced yet. The current implementation is based on Taming-3DGS, gsplat and Gaussianimage, and more implementation details will be open sourced step by step in the future.

Requirements

Follow the instructions in the Taming 3DGS repository and GaussianImage repository to setup the environment. Or simply run the following commands:

cd src/gimage/gaussianimage/gsplat && python setup.py build_ext --inplace -j32
cp build/*/*/csrc.so gsplat/

git clone https://github.com/humansensinglab/taming-3dgs.git --recursive
cd taming-3dgs
pip install submodules/*

Dataset and Checkpoints

We use the following datasets for evaluation:

NeRF On-the-go dataset

RobustNeRF dataset

Please modify the DATA_DIR in render_and_metric.sh to the path of your dataset.

We provide the checkpoints for rendering and evaluation in HybridGS checkpoints.

Please modify the CKPT_PATH in render_and_metric.sh to the path of your checkpoints.

Usage

./render_and_metric.sh

Citation

If you find our code or paper useful, please cite:

@InProceedings{lin2024hybridgs,
    author = {Jingyu Lin, Jiaqi Gu, Lubin Fan, Bojian Wu, Yujing Lou, Renjie Chen, Ligang Liu, Jieping Ye},
    title = {HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting},
    booktitle = {Arxiv},
    year = {2024}
}