GauSS-MI:
Gaussian Splatting Shannon Mutual Information for Active 3D Reconstruction

RSS 2025
Yuhan Xie, Yixi Cai, Yinqiang Zhang, Lei Yang, and Jia Pan
Paper | arXiv | YouTube | bilibili

📢 News

• 15 Jul. 2025: 🐳 Docker image released.
• 26 Jun. 2025: GauSS-MI Code Release! 🔓
• 28 May. 2025: Submodule Differential Gaussian Rasterization with GauSS-MI released.

💡 Highlights

GauSS-MI is a real-time active view selection metric for 3D Gaussian Splatting, optimizing high-visual-quality 3D reconstruction.

🔗 BibTeX

If you find our code/work useful, please consider citing our work. Thank you!

@INPROCEEDINGS{YuhanRSS25, 
    AUTHOR    = {Yuhan Xie AND Yixi Cai AND Yinqiang Zhang AND Lei Yang AND Jia Pan}, 
    TITLE     = {{GauSS-MI: Gaussian Splatting Shannon Mutual Information for Active 3D Reconstruction}}, 
    BOOKTITLE = {Proceedings of Robotics: Science and Systems}, 
    YEAR      = {2025}, 
    ADDRESS   = {LosAngeles, CA, USA}, 
    MONTH     = {June}, 
    DOI       = {10.15607/RSS.2025.XXI.030} 
} 

Acknowledgements

This project builds heavily on MonoGS and 3D Gaussian Splatting. We thank the authors for their excellent work! If you use our code, please consider citing the papers as well.

⚙️ Environment Setup

0. Requirements

• Most hardware and software requirements same as 3D Gaussian Splatting Optimizer.
• Conda (recommended for easy setup)
• CUDA Toolkit 11.8
• ROS1 Noetic + Ubuntu 20.04

1. Clone the Repository

mkdir -p ~/ws_gaussmi/src && cd ~/ws_gaussmi/src    # ROS Workspace
git clone https://github.com/JohannaXie/GauSS-MI.git --recursive    # https
# git clone git@github.com:JohannaXie/GauSS-MI.git --recursive        # ssh

2. Conda Setup

cd GauSS-MI
conda env create -f environment.yml
conda activate GauSS-MI

3. Declare the Python path under your conda environment on the first line of scripts/gs_map.py, which could be #!/opt/conda/envs/GauSS-MI/bin/python or #!/home/{YourUserName}/anaconda3/envs/GauSS-MI/bin/python.

4. ROS Setup

cd ~/ws_gaussmi
catkin build -DPYTHON_EXECUTABLE=/opt/conda/envs/GauSS-MI/bin/python    # Your python path under conda

🐳 Quick Setup with Docker

1. Pull the docker image and build the container

# Pull Image
docker pull johanna17/gauss-mi:v1
# Run the Docker container
docker run -it -d --gpus all -e DISPLAY \
    -v /tmp/.X11-unix:/tmp/.X11-unix \
    -e QT_X11_NO_MITSHM=1 \
    -e NVIDIA_DRIVER_CAPABILITIES=all \
    --shm-size=32G \
    --ipc=host \
    --network=host \
    --cap-add SYS_PTRACE \
    --cap-add SYS_ADMIN \
    --privileged \
    -v /dev:/dev \
    --name=GauSS-MI \
    -v $HOME/docker_envs/GauSS-MI:/home/do \
    -u do \
    -w /home/do \
    johanna17/gauss-mi:v1
# Go into the Container
docker exec -it GauSS-MI bash
# docker exec -itd GauSS-MI terminator  # Or use terminator

2. [In Container] Activate ROS and Conda

source /opt/ros/noetic/setup.zsh
source /opt/conda/bin/activate 
/opt/conda/bin/conda init zsh

3. [In Container] Clone the Repository and build

# Clone the repo
mkdir -p ~/ws_gaussmi/src && cd ~/ws_gaussmi/src    # ROS Workspace
git clone https://github.com/JohannaXie/GauSS-MI.git --recursive    # https
# Build the ros package
cd ~/ws_gaussmi
catkin config --extend /opt/ros/noetic
catkin build -DPYTHON_EXECUTABLE=/opt/conda/envs/GauSS-MI/bin/python

🚀 Quick Start

Use example rosbag

• Download the example robag here.
• Launch GauSS-MI

  conda activate GauSS-MI
  cd ~/ws_gaussmi
  source devel/setup.zsh
  roslaunch gs_mapping gaussmi_rosbag_oildrum.launch

• Run rosbag in another terminal

  # source /opt/ros/noetic/setup.zsh
  rosbag play GauSS-MI_example1_oildrum.bag

🦾 Implement on Your Robot

ROS Publisher & Subscriber

• Publish your collected RGB image, depth image, and camera pose at initialization viewpoints and the next-best-view on rostopic, respectively:

  /camera/bgr
  /camera/depth
  /camera/pose
  

• Subscribe to the next-best-view pose on rostopic:

  /gaussmi/nbv_pose
  

Simulation or Real-World Robot Implementation

To implement our code on your robot with the motion planner and controller. You may need to customize the code in scripts/active_recon/active_manage.py. Specifically:

• Robot Initialization
• Collision Check
• Robot Motion Planner & Waiting for Robot's Navigation.

Customize Parameters

You may need to customize the parameters for your reconstruction environment in the configuration file configs/eg_data_config/active.yaml.

Launch GauSS-MI and Your Robot

• Launch GauSS-MI

  conda activate GauSS-MI
  cd ~/ws_gaussmi
  source devel/setup.zsh
  roslaunch gs_mapping gaussmi_active.launch

• Launch Your Robot