OpenGS-Fusion: Open-Vocabulary Dense Mapping with Hybrid 3D Gaussian Splatting for Refined Object-Level Understanding

Dianyi Yang, Xihan Wang, Yu Gao, Shiyang Liu, Bohan Ren, Yufeng Yue, Yi Yang*

IROS 2025

Project | Video

This repository is intended to provide an engineering implementation of our paper, and we hope it will contribute to the community. If you have any questions, feel free to contact us.

Install requirements

conda create -n opengsfusion python==3.9
conda activate opengsfusion
conda install pytorch==2.0.0 torchvision==0.15.0 torchaudio==2.0.0 pytorch-cuda=11.8 -c pytorch -c nvidia
conda install cmake
pip install -r requirements.txt

Also, PCL is needed for fast-gicp submodule.

Install submodules

conda activate opengsfusion
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn
pip install submodules/MobileSAM

export OPENGS_ENV=/path/to/your/anaconda3/envs/opengsfusion
pip install submodules/vdbfusion

cd submodules/fast_gicp
mkdir build
cd build
cmake ..
make
cd ..
python setup.py install --user

Install mobilesam2 weights

Please download the MobileSAMv2 weights from the following link Driver After downloading, place the files in opengs_fusion/submodules/MobileSAM/MobileSAMv2/weight. The directory structure should look like:

opengs_fusion/submodules/MobileSAM/MobileSAMv2/weight
├── l2.pt
├── mobile_sam.pt
├── ObjectAwareModel.pt

• Replica

  • Download

    bash download_replica.sh

  • Configure

    Please modify the directory structure to ours.

    The original structure

    Replica
        - room0
            - results (contain rgbd images)
                - frame000000.jpg
                - depth000000.jpg
                ...
            - traj.txt
        ...

    Our structure

    Replica
        - room0
            - images (contain rgb images)
                - frame000000.jpg
                ...
            - depth_images (contain depth images)
                - depth000000.jpg
                ...
            - traj.txt
        ...

• Scannet

  • Download follow scannet

    Our structure

    data
        - scene0046_00
            - rgb (contain rgb images)
                - 0.png
                ...
            - depth (contain depth images)
                - 0.png
                ...
            - pose (contain poses)
                - 0.txt
                ...
            - traj.txt
        ...

    The traj.txt file here is generated by running ./datasets_process/convert_pose_2_traj.py.

• Custom datasets:

  For custom datasets, you should format your data to match either the Replica or ScanNet dataset structures. Additionally, you'll need to create a camera configuration file specifying your camera's intrinsic parameters config.txt:

  ## camera parameters
  W H fx fy cx cy depth_scale depth_trunc dataset_type
  640 480 577.590698 578.729797 318.905426 242.683609 1000.0 5.0 scannet

  You can put this config in ./config directory.

• Replica

  bash ./bash/train_replica_with_sem.sh

• Scannet

  bash ./bash/train_scannet_with_sem.sh

The pipeline has two steps for each dataset:

• Feature Extraction: Runs mobilesamv2_clip.py to extract 2D SAM masks and CLIP features.

  python mobilesamv2_clip.py --image_folder /path/to/images --output_dir /path/to/output --save_results 

• 3D Mapping: Runs opengs_fusion.py to build semantic GS maps.

  python opengs_fusion.py --dataset_path /path/to/dataset --config /path/to/config.txt --output_path /path/to/output --rerun_viewer --save_results

We also put some potential bugs to issue, please check it out~.

After completing the mapping process, you can visualize and interact with the semantic maps using the following commands:

python show_lang_embed.py \
    --dataset_path /path_to_replica/office0 \
    --config ./configs/Replica/caminfo.txt \
    --scene_npz /path_to_replica_output/office0/office0_default_each/gs_scene.npz \
    --dataset_type replica \
    --view_scale 2.0

python show_lang_embed.py \
    --dataset_path /path_to_scannet/scene0062_00 \
    --config ./configs/Scannet/scene0062_00.txt \
    --scene_npz /path_to_scannet_output/scene0062_00/default_with_sem/gs_scene.npz \
    --dataset_type scannet \
    --view_scale 3.0

Here, users can freely adjust the viewing angle in the interface. We also provide a text box for real-time querying and threshold adjustment. All tests were conducted on an Ubuntu system with a 2K resolution.

• T: Toggle between color and label display modes.
• J: Highlight selected object.
• K: Capture screenshot of current view.
• O: Print current view information.
• M: Switch between different camera views.
• P: Downsample the point cloud.
• =: Save current mask point cloud.
• L: Toggle voxel visualization.

Rerun viewer shows the means of trackable Gaussians, and rendered image from reconstructed 3dgs map.

The demo show here is supported by GS_ICP_SLAM.

You just need to add --rerun_viewer to the command when running opengs_fusion.py. For example:

python opengs_fusion.py --dataset_path /path/to/dataset --config /path/to/config.txt --output_path /path/to/output --rerun_viewer

This work builds upon the following outstanding open-source projects:

• GS_ICP_SLAM - For their foundational work on Gaussian Splatting with ICP-based SLAM
• VDBFusion - For their efficient volumetric mapping framework

We're deeply grateful to the researchers behind these projects for sharing their work with the community.

If you find this work useful for your research, please cite our paper:

@inproceedings{yang2025opengs-fusion,

}