Autonomous Learning Group University of Tübingen

Combinatorial Optimization

Deep Graph Matching

figure image

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.

Publications

Deep Graph Matching via Blackbox Differentiation of Combinatorial Solvers
Deep Graph Matching via Blackbox Differentiation of Combinatorial Solvers
Michal Rolinek, Paul Swoboda, Dominik Zietlow, Anselm Paulus, Vit Musil, Georg Martius
Computer Vision – ECCV 2020  ·  2020
arXiv