Computer Vision News - July 2020

57 Guillem Brasó on integrating learning into the overall data association task, including the partitioning algorithm. We hope that the community will realize that graphs are still a great modeling tool for MOT in the deep learning era: they can be integrated within end-to- end pipelines and they don’t necessarily require expensive optimization to yield great results. With the help of message passing networks, classical graph methods can benefit from the full power of learning. Overview of the method, from detections to graph construction: the tracks are found using message passing networks. Best of CVPR 2020 combinatorial optimization algorithms. In summary, we propose a fully differentiablemodel for data association that can replace expensive optimization algorithms while running at a fraction of their cost and yielding significantly better results . Where do we go from here? We expect our approach to pave the way for future work to go beyond feature extraction for MOT and focus, instead, "We replace combinatorial optimization solvers by a message passing network that can directly predict trajectories from a set of detections by operating on the graph domain"