image-match/python/MambaGlue

MambaGlue 🐍 @ICRA2025
Fast and Robust Local Feature Matching With Mamba

Kihwan Ryoo · Hyungtae Lim · Hyun Myung

<img src="assets/Visualization.png" alt="example" width=40%><img src=assets/demo_sacre.gif alt="animated" width="40%"/></a>
<br>
<em>MambaGlue is a hybrid neural network combining the Mamba and the Transformer architectures to match local features.<br></em>

MambaGlue :snake:

Main branch includes the standard MambaGlue model. Thanks to CVG Lab, you can easily train and evaluate the model and visualize the results on glue-factory branch and hloc branch.

:dart: Training and Evaluation (glue-factory branch)

Using Glue Factory, set MambaGlue for a matcher model and train MambaGlue with any local features on your own or open-sourced dataset! It will take about 1 week for one trial. Additionally, you can evaluate its performance compared with other baseline models on benchmarks such as HPatches and MegaDepth.

:magic_wand: Visualization and Evaluation (hloc branch)

Using Hierarchical-Localization, set MambaGlue for a matcher model and run MambaGlue for Structure-from-Motion and visual localization!

:desktop_computer: Tested Environment

• Linux (UBUNTU 20.04)
• NVIDIA GPU (TITAN V || RTX 3080 || other Ampere architectures)
• CUDA 11.8
• CUDNN 8
• PyTorch 2.1.0
• Python 3.8

:keyboard: Install

Install MambaGlue:

git clone https://github.com/state-spaces/mamba && cd mamba
pip install .
cd ..
git clone https://github.com/url-kaist/MambaGlue.git && cd MambaGlue
python -m pip install -e .

You can set up the environment starting from our docker image or PyTorch official docker image.

:clipboard: To Do

• Release demo code
• Update branches
• ONNX

:memo: Citation

@article{ryoo2025mambaglue,
  title={{MambaGlue: Fast and Robust Local Feature Matching With Mamba}},
  author={Ryoo, Kihwan and
          Lim, Hyungtae and
          Myung, Hyun},
  journal={arXiv preprint arXiv:2502.00462},
  year={2025}
}

License

The MambaGlue code provided in this repository is released under the Apache-2.0 license.