Edstedt Johan ; Sun Qiyu ; Bökman Georg ; Wadenbäck Mårten ; Felsberg Michael

Abstract

Feature matching is an important computer vision task that involvesestimating correspondences between two images of a 3D scene, and dense methodsestimate all such correspondences. The aim is to learn a robust model, i.e., amodel able to match under challenging real-world changes. In this work, wepropose such a model, leveraging frozen pretrained features from the foundationmodel DINOv2. Although these features are significantly more robust than localfeatures trained from scratch, they are inherently coarse. We therefore combinethem with specialized ConvNet fine features, creating a precisely localizablefeature pyramid. To further improve robustness, we propose a tailoredtransformer match decoder that predicts anchor probabilities, which enables itto express multimodality. Finally, we propose an improved loss formulationthrough regression-by-classification with subsequent robust regression. Weconduct a comprehensive set of experiments that show that our method, RoMa,achieves significant gains, setting a new state-of-the-art. In particular, weachieve a 36% improvement on the extremely challenging WxBS benchmark. Code isprovided at https://github.com/Parskatt/RoMa