Voxelized 3D Gaussian Rrepresentation for Dense Visual SLAM on Embedded Vision System

IJCV 2025

Table of Contents

1. Installation
2. Online Demo
3. Usage
4. Downloads
5. Benchmarking
6. Acknowledgement
7. Citation

Notes

Our method is a plug-and-play approach that can be integrated with different GS-SLAM frameworks. We will maintain separate branches for different versions of the method. The current version supports MonoGS and Gaussian-SLAM.

• We have updated the README.md and are preparing to open-source our code！
• Code for main parts, including optimizer, renderer, tracking and mapping modules, Voxelized Scene Representation
• Installation setup

🛠️ Setup

The code has been tested on:

• Ubuntu 22.04 LTS, Python 3.10.14, CUDA 12.2, GeForce RTX 4090/RTX 3090
• CentOS Linux 7, Python 3.12.1, CUDA 12.4, A100/A6000

📦 Repository

Clone the repo with --recursive because we have submodules:

git clone --recursive git@github.com:dtc111111/Compact_GSSLAM.git
cd VCGS-SLAM

💻 Installation

Make sure that gcc and g++ paths on your system are exported:

export CC=<gcc path>
export CXX=<g++ path>

To find the gcc path and g++ path on your machine, you can use which gcc.

Then setup environment from the provided conda environment file,

conda create -n vcgs-slam -c nvidia/label/cuda-12.1.0 cuda=12.1 cuda-toolkit=12.1 cuda-nvcc=12.1
conda env update --file environment.yml --prune
conda activate vcgs-slam
pip install -r requirements.txt

You will also need to install hloc for loop detection and 3DGS registration.

cd thirdparty/Hierarchical-Localization
python -m pip install -e .
cd ../..

We tested our code on RTX4090 and RTX A6000 GPUs respectively and Ubuntu22 and CentOS7.5.

🚀 Usage

Here we elaborate on how to load the necessary data, configure Gaussian-SLAM for your use-case, debug it, and how to reproduce the results mentioned in the paper.

Downloading the Datasets

We tested our code on Replica, TUM_RGBD, ScanNet, and ScanNet++ datasets. We also provide scripts for downloading Replica and TUM_RGBD in scripts folder. Install git lfs before using the scripts by running git lfs install.

For reconstruction evaluation on Replica, we follow Co-SLAM mesh culling protocal, please use their code to process the mesh first.

For downloading ScanNet, follow the procedure described on here.
Pay attention! There are some frames in ScanNet with inf poses, we filter them out using the jupyter notebook scripts/scannet_preprocess.ipynb. Please change the path to your ScanNet data and run the cells.

For downloading ScanNet++, follow the procedure described on here.

The config files are named after the sequences that we used for our method.

Running the code

Start the system with the command:

python run_slam.py configs/<dataset_name>/<config_name> --input_path <path_to_the_scene> --output_path <output_path>

You can also configure input and output paths in the config yaml file.

Reproducing Results

You can reproduce the results for a single scene by running:

python run_slam.py configs/<dataset_name>/<config_name> --input_path <path_to_the_scene> --output_path <output_path>

If you are running on a SLURM cluster, you can reproduce the results for all scenes in a dataset by running the script:

./scripts/reproduce_sbatch.sh

Please note the evaluation of depth_L1 metric requires reconstruction of the mesh, which in turns requires headless installation of open3d if you are running on a cluster.

✏️ Acknowledgement

Our implementation is heavily based on Gaussian-SLAM and MonoGS and Loop-Splat. We thank the authors for their open-source contributions. If you use the code that is based on their contribution, please cite them as well. We thank Yue Pan for the fruitful discussion.