GeoMoE: Divide-and-Conquer Motion Field Modeling with Mixture-of-Experts for Two-View Geometry

Jiajun Le · Jiayi Ma^* ·

Wuhan University    

^*Corresponding author

GeoMoE models two-view matching as motion-field consensus and explicitly decomposes the global motion field into multiple sub-fields, each handled by a specialized Mixture-of-Experts module guided by inlier-prior cues. This decoupling reduces interference across motion regimes, yielding more reliable inlier prediction.

Requirements

Please use Python 3.8.5, opencv-contrib-python (4.12.0.88) and Pytorch (>= 1.9.1). Other dependencies should be easily installed through pip or conda.

Example scripts

Run the demo

You can run the feature matching for two images with GeoMoE.

cd demo && python demo.py

Datasets and Pretrained models

Download the pretrained models from here.

Download YFCC100M dataset.

bash download_data.sh raw_data raw_data_yfcc.tar.gz 0 8
tar -xvf raw_data_yfcc.tar.gz

Download SUN3D testing (1.1G) and training (31G) dataset if you need.

bash download_data.sh raw_sun3d_test raw_sun3d_test.tar.gz 0 2
tar -xvf raw_sun3d_test.tar.gz
bash download_data.sh raw_sun3d_train raw_sun3d_train.tar.gz 0 63
tar -xvf raw_sun3d_train.tar.gz

Test pretrained model

We provide the models trained on the YFCC100M and SUN3D datasets, as detailed in our AAAI paper. By running the test script, you can obtain results similar to those presented in our paper. Note that the generated putative matches may differ if the data is regenerated.

cd ../test 
python test.py

To adjust the default settings for test, you can edit the ../test/config.py.

Train model on YFCC100M or SUN3D

After generating dataset for YFCC100M, run the tranining script.

cd ../core 
python main.py

Citation

If you find this project useful, please cite:

@inproceedings{Le2026GeoMoE,
  title={GeoMoE: Divide-and-Conquer Motion Field Modeling with Mixture-of-Experts for Two-View Geometry},
  author={Le, Jiajun and Ma, Jiayi},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2026}
}

Acknowledgement

Parts of this code are adapted from DeMo and DeMatch. If you use any of the DeMo/DeMatch-derived code, please cite their papers.

@inproceedings{lu2025demo,
  title={Deep Motion Field Consensus with Learnable Kernels for Two-view Correspondence Learning},
  author={Lu, Yifan and Le, Jiajun and Li, Zizhuo and Yuan, Yixuan and Ma, Jiayi},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={6},
  pages={5829--5837},
  year={2025}
}
@inproceedings{zhang2024dematch,
  title={Dematch: Deep decomposition of motion field for two-view correspondence learning},
  author={Zhang, Shihua and Li, Zizhuo and Gao, Yuan and Ma, Jiayi},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={20278--20287},
  year={2024}
}