Correspondence-Aware Image-to-Image Translation Unlocks Cross-Modal Matching via Single-Modality Priors
"Patch Your Matcher" addresses the cross-modal part of the registration problem. The method learns image-to-image translations that make images from different modalities easier to match with existing image matchers.
Instead of training only for visual realism, the method trains translations for geometric correspondence. This allows strong frozen matchers such as ELoFTR, LightGlue and RoMA to operate on cross-modal image pairs.
Find out more:
- Project page: https://xaf-cv.github.io/pym/
- Code: https://github.com/xaf-cv/patch-your-matcher
Citation:
A. Frolov and V. Rodehorst, ‘Patch Your Matcher: Correspondence-Aware Image-to-Image Translation Unlocks Cross-Modal Matching via Single-Modality Priors’, in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2026, pp. 7913–7924.