Computer Vision News - July 2020

55 Best of CVPR 2020 Guillem Brasó Learning a Neural Solver for Multiple Object Tracking Guillem Brasó is an MSc. student in mathematics and a research assistant under the supervision of Laura Leal- Taixé at the Dynamic Vision and Learning Group of the Technical University of Munich. TLDR: In this work, we revisit the classical min-cost flow graph formulation of multi-object tracking (MOT) from a deep learning perspective . 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. by Guillem Brasó Guillem, Laura and other members of the Dynamic Vision and Learning group (Ismail Elezi, Maxim Maximov, Aljosa Osep, and Yihong Xu) after the CVPR submission all-nighter (deadline is 9am in Europe!). How to work in the MOT domain? Graphs have typically been the standard modeling tool for MOT. Object detections are seen as nodes, and possible connections forming trajectories are modeled as edges. In this setting, obtaining object trajectories (i.e. data association) is done by partitioning the graph via combinatorial optimization . Typically, top-performing methods rely on particularly expensive optimization frameworks, and hence, graph-based MOT has typically suffered from a severe speed/accuracy tradeoff. Moreover, learning within this graph- based formulation is not trivial for