Establishing robust and accurate correspondences is a fundamental backbone to\nmany computer vision algorithms. While recent learning-based feature matching\nmethods have shown promising results in providing robust correspondences under\nchallenging conditions, they are often limited in terms of precision. In this\npaper, we introduce S2DNet, a novel feature matching pipeline, designed and\ntrained to efficiently establish both robust and accurate correspondences. By\nleveraging a sparse-to-dense matching paradigm, we cast the correspondence\nlearning problem as a supervised classification task to learn to output highly\npeaked correspondence maps. We show that S2DNet achieves state-of-the-art\nresults on the HPatches benchmark, as well as on several long-term visual\nlocalization datasets.\n