Staying in the research area of computer vision, we tackle the problem of keypoint matching. Leveraging our blackbox-backprop framework, we carefully extract keypoint descriptors using visual and geometrical information and use a state-of-the-art solver to solve a bipartite graph matching problem. The resulting end-to-end trainable architecture set the state-of-the-art on challenging keypoint matching benchmarks when it was published at ECCV 2020.
