Abstract
Establishing robust and accurate correspondences is a fundamental backbone tomany computer vision algorithms. While recent learning-based feature matchingmethods have shown promising results in providing robust correspondences underchallenging conditions, they are often limited in terms of precision. In thispaper, we introduce S2DNet, a novel feature matching pipeline, designed andtrained to efficiently establish both robust and accurate correspondences. Byleveraging a sparse-to-dense matching paradigm, we cast the correspondencelearning problem as a supervised classification task to learn to output highlypeaked correspondence maps. We show that S2DNet achieves state-of-the-artresults on the HPatches benchmark, as well as on several long-term visuallocalization datasets.