David Nordström

It is found that, despite performing poorly in zero-shot matching, especially in later layers, feed-forward reconstruction models provide strong representations for linear probing and full matching pipelines, and their raw predictions alone enable competitive matching, albeit only under moderate viewpoint changes and modality gaps.

Surprisingly, learning rotation invariance directly in the feature descriptor yields similar performance to handling it in the matcher, but with improved speed and no loss in upright accuracy.

LoMa shatters performance ceilings in local feature matching, achieving state-of-the-art results by scaling data, model size, and compute, proving that even well-trodden areas can yield surprising gains with modern techniques.