Update

March 13, 2026 · View on GitHub

CG-SLAM: Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field
[ECCV 2024]

Jiarui Hu1† · Xianhao Chen1† · Boyin Feng1 · Guanglin Li1 · Liangjing Yang2
Hujun Bao1 · Guofeng Zhang1 · Zhaopeng Cui1*
1 State Key Lab of CAD&CG, Zhejiang University
2 ZJU-UIUC Institute, International Campus, Zhejiang University
* Corresponding author. † Equal contribution.

Project Page | Paper | Video

This repository contains the official implementation of CG-SLAM, a dense RGB-D SLAM system built on a consistent uncertainty-aware 3D Gaussian field. The project targets high-quality tracking, mapping, and rendering while keeping the system efficient enough for practical scene reconstruction experiments.

CG-SLAM teaser

CG-SLAM visualization

Table of Contents
  1. Update
  2. Submodule
  3. Installation
  4. Usage
    1. Run
    2. Evaluation
  5. Acknowledgement
  6. Citation

Update

  • Code for Diff-rasterization(w/pose --> 4✖️4 Transformation Matrix T)
  • Our paper is accepted by ECCV 2024, and our code is coming soon!!!
  • Code for RGBD-SLAM
  • Code for Evaluation

Submodule

We have proposed a comprehensive mathematical theory on derivatives w.r.t. pose in 3D Gaussian splatting framework. Additionally, we have developed a specialized CUDA framework tailored for the SLAM task, decoupling the tracking and mapping components. For more details, please refer to the provided diff-gaussian-rasterization.

Installation

Requirements

  • Linux
  • Python 3.8
  • CUDA-capable NVIDIA GPU
  • PyTorch with a CUDA build compatible with your local toolkit
git clone <your-release-url> CG-SLAM
cd CG-SLAM
git submodule update --init --recursive

conda create -n cg-slam python=3.8
conda activate cg-slam

Install PyTorch first. Pick the wheel that matches your CUDA environment. For example:

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118

Install the remaining Python dependencies and local extensions:

pip install pyyaml scipy scikit-image opencv-python plyfile open3d tensorboard mathutils==2.81.2 tqdm trimesh imageio
pip install -e ./submodules/diff-gaussian-rasterization
pip install -e ./submodules/simple-knn

torch-scatter should be installed with a wheel that matches your PyTorch and CUDA versions. If it is not already available in your environment, install the matching build before running CG-SLAM.

Usage

Run

The provided YAML files under configs/ are examples for different datasets. Before running, either:

  • edit data.input_folder and data.output_folder in the selected config file, or
  • override them from the command line with --input_folder and --output_folder

Replica

python run.py \
  --config ./configs/Replica/office0.yaml \

TUM RGB-D

python run.py \
  --config ./configs/TUM/fr1_desk.yaml \

ScanNet

python run.py \
  --config ./configs/ScanNet/scene0000.yaml \

Evaluation

ATE

For Replica scenes, you can evaluate the estimated trajectory with:

python tools/eval_ate.py --config ./configs/Replica/office0.yaml

For TUM RGB-D and ScanNet, the current script expects the corresponding ground-truth pose file to be available in the output directory as gt_pose.txt.

Acknowledgement

We sincerely thank the authors of the following projects:

Their excellent open-source work made this project possible.

Citation

@article{hu2024cg,
  title={CG-SLAM: Efficient Dense RGB-D SLAM in a Consistent Uncertainty-aware 3D Gaussian Field},
  author={Hu, Jiarui and Chen, Xianhao and Feng, Boyin and Li, Guanglin and Yang, Liangjing and Bao, Hujun and Zhang, Guofeng and Cui, Zhaopeng},
  journal={arXiv preprint arXiv:2403.16095},
  year={2024}
}