Large-Scale Gaussian Splatting SLAM

This is an official implementation of our work published in ICRA'25. Project Page

Large-Scale Gaussian Splatting SLAM

Zhe Xin¹, Chenyang Wu^{1, 2}, Penghui Huang¹, Yanyong Zhang², Yinian Mao¹, and Guoquan Huang^{1, 3
1}Meituan UAV, Beijing, China, ²School of Computer Science and Technology, University of Science and Technology of China, Hefei, China, ³Dept. of Mechanical Engineering, Computer and Information Sciences, University of Delaware, Newark, DE, USA

Paper (arXiv)

Installation

Please follow the instructions below to install the repo and dependencies.

git clone https://github.com/lsg-slam/LSG-SLAM.git
cd LSG-SLAM

Install the environment

# Create conda environment
conda create -n lsgslam python=3.10
conda activate lsgslam

# Install the requirements
conda install -c "nvidia/label/cuda-11.6.0" cuda-toolkit
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge
pip install -r requirements.txt

# Build extension 
cd diff-gaussian-rasterization-w-depth.git
python setup.py install
pip install .

Dataset

We use EuRoC and KITTI datasets.

Run

Before run LSG-SLAM, you need to run tools/euroc_parser/operate_euroc_data.py and tools/kitti_parser/operate_kitti_data.py first to get depth images and global features.

Run scripts/loop_closure.py to run front end and loop closure:

python scripts/loop_closure.py configs/euroc/lsgslam.py

Run tools/loop_closure/pose_graph_part_optim.py to run back end (pose graph and structure refine):

python tools/loop_closure/pose_graph_part_optim.py

Acknowledgement

Our codebase builds on the code in SplaTAM.

Citation

If you find our code or paper useful for your research, please consider citing:

@article{xin2025large,
  title={Large-Scale Gaussian Splatting SLAM},
  author={Xin, Zhe and Wu, Chenyang and Huang, Penghui and Zhang, Yanyong and Mao, Yinian and Huang, Guoquan},
  journal={arXiv preprint arXiv:2505.09915},
  year={2025}
}