JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba

This is a PyTorch implementation of JamMa for CVPR'25 paper “JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba”.

JamMa is a lightweight image matcher that enables fast internal and mutual interaction of images with joint Mamba.

Installation

conda env create -f environment.yaml
conda activate jamma
pip install -r requirements.txt 
pip install mamba-ssm==2.0.3  # require CUDA>11.6, recommend installing from .whl file (way faster)

Demo

We provide a simple demo for matching image pairs and generating visualizations

cd demo
python demo.py #[--image1 path1 --image2 path2]

Reproduce the testing results

The test and training can be downloaded by download link provided by LoFTR.

You need to setup the testing subsets of ScanNet and MegaDepth first. We create symlinks from the previously downloaded datasets to data/{{dataset}}/test.

# set up symlinks
ln -s /path/to/scannet-1500-testset/* /path/to/project/data/scannet/test
ln -s /path/to/megadepth-1500-testset/* /path/to/project/data/megadepth/test

bash scripts/reproduce_test/outdoor.sh

You should expect the following results

{'auc@10': 0.7738815790570285,
 'auc@20': 0.8651362212886763,
 'auc@5': 0.641769096495131,
 'num_matches@1e-04': 4021.4746666666665,
 'prec@1e-04': 0.8891183971633494}

Training on MegaDepth

Please follow the training doc for data organization

bash scripts/reproduce_train/outdoor.sh

Part of the code is based on the LoFTR and XoFTR repository under their own separate license, terms and conditions.

We are grateful to the authors for their contribution of the well-structured code.

If you find this project useful, please cite:

@misc{lu2025jamma,
    title={JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba},
    author={Xiaoyong Lu and Songlin Du},
    year={2025},
    eprint={2503.03437},
    archivePrefix={arXiv},
    url={https://arxiv.org/abs/2503.03437},
}