Dong-Uk Seo · Jinwoo Jeon · Eungchang Mason Lee · Hyun Myung

Given a monocular video sequence, GaussianFlow SLAM tracks the camera trajectory and reconstructs a photorealistic 3DGS map by leveraging optical flow as geometric supervision, without any depth measurements.

git clone --recursive https://github.com/url-kaist/gaussianflow-slam.git
cd gaussianflow-slam

System packages:

sudo apt install build-essential libeigen3-dev libopencv-dev python3-tk \
    libgl1 libglib2.0-0 libsm6 libxext6 libxrender1

Python environment:

python3 -m venv .venv
source .venv/bin/activate

pip install -r requirements.txt

Adjust the torch, torchvision, torchaudio versions and the --extra-index-url in requirements.txt to match your CUDA version.

CUDA extensions (built locally from thirdparty/ and src/):

pip install thirdparty/lietorch
pip install thirdparty/pytorch_scatter
pip install thirdparty/simple-knn
pip install -e thirdparty/diff-gaussian-rasterization \
    --use-pep517 --no-build-isolation --force-reinstall
pip install -e .

Pre-trained weights:

./tools/download_model.sh

Reference environment: Ubuntu 20.04, Python 3.10, CUDA 11.8, PyTorch 2.1.2+cu118.

Download the ASL-format sequences from the EuRoC MAV page and convert the ground truth to TUM format (gt_converted/<seq>_gt.txt).

./tools/download_tum.sh

./tools/run_gaussianflow_slam.sh V101          # GUI
./tools/run_gaussianflow_slam_nogui.sh V101    # headless

./tools/run_gaussianflow_mapping_only.sh V101          # GUI
./tools/run_gaussianflow_mapping_only_nogui.sh V101    # headless

Sequences: V101, MH01, fr1/desk, fr2, fr3. Options and environment variables: tools/run_gaussianflow_modes.sh --help.

Edit DATA_PATH at the top of each script to point to your local dataset, then:

./tools/run_euroc_gsflow.sh     # EuRoC
./tools/run_tum_gsflow.sh       # TUM-RGBD

Extra demo.py flags can be appended to the script call, e.g.:

./tools/run_euroc_gsflow.sh --t1=600 --stride=2

You only need (1) a folder of RGB frames named so they sort in capture order and (2) a one-line calibration file fx fy cx cy [k1 k2 p1 p2 [k3 [k4 k5 k6]]]. Then:

python3 demo.py --imagedir=<path/to/rgb> --calib=<calib.txt> \
    --gsconfig_path=configs/mono_slam.yaml --upsample --disable_vis \
    --output=results/<run>/

Passing --gt_path makes demo.py report ATE and write traj_full.txt / traj_kf.txt to the output directory. To recompute ATE from a finished run:

python3 tools/compute_ate.py results/<run>/traj_full.txt <gt.txt>

This work builds on many open-source projects. We extend our gratitude to the authors:

• DROID-SLAM
• 3D Gaussian Splatting
• MonoGS
• GaussianFlow

GaussianFlow SLAM is released under the GPL-3.0 License. For a list of code dependencies which are not property of the authors of GaussianFlow SLAM, please check Dependencies.md.

If you found this work useful in your own research, please consider citing:

@article{seo2026gaussianflow,
  title={GaussianFlow SLAM: Monocular Gaussian Splatting SLAM Guided by GaussianFlow},
  author={Seo, Dong-Uk and Jeon, Jinwoo and Lee, Eungchang Mason and Myung, Hyun},
  journal={IEEE Robotics and Automation Letters},
  year={2026},
  publisher={IEEE}
}