README.md

AerialExtreMatch: A Benchmark for Extreme-View Image Matching and Localization

Project Page | Paper

AerialExtreMatch: A Benchmark for Extreme-View Image Matching and Localization
Rouwan Wu¹, Zhe Huang², Xingyi He², Yan Liu³, Shen Yan¹, Sida Peng², Maojun Zhang^1†, Xiaowei Zhou^2†
1NUDT, ²State Key Lab of CAD&CG, ZJU, ³HUST

2025

We introduce AerialExtreMatch, a large-scale, high-fidelity benchmark tailored for extreme-view image matching and UAV localization. It consists of three datasets: Train Pair, Evaluation Pair, and Localization. All code and datasets are readily available for public access.

Resources

Important

In our paper, TWO seperate codebases are provided: benchmarking and code of our pretrained RoMa model.
To increase simplicity and consistency, we slightly abuse the concept of git branches and make the two codebases as branches of this repository.

• Code
  • Benchmark branch: source code for the benchmark, including feature matching and localization pipelines for models mentioned in the paper.
  • RoMa branch: the code we use to train our RoMa model.
• Dataset
  • AerialExtreMatch-Train: corresponds to Train Pair set.
  • AerialExtreMatch-Benchmark: corresponds to Evaluation Pair set.
  • AerialExtreMatch-Localization: corresponds to Localization set.
• Checkpoints: see [Release].

License

MIT License

Citation

If you find our work useful, please consider citing:

@article{wu2026aerialextrematch,
  title   = {AerialExtreMatch: A Benchmark for Extreme-View Image Matching and Localization},
  author  = {Wu, Rouwan and Huang, Zhe and He, Xingyi and Liu, Yan and Yan, Shen and Peng, Sida and Zhang, Maojun and Zhou, Xiaowei},
  journal = {IEEE Robotics and Automation Letters},
  year    = {2026},
  doi     = {10.1109/LRA.2026.3673915}
}