Graph Neural Networks for Multiple Object Tracking

Multiple object tracking(MOT) is the task of studying object appearance and movements to analyze their trajectories. For a given input video the algorithm is supposed to output which portions of the image represent the same object in different frames of the video. Algorithms like these can be used to solve some exciting problems like analyzing a particular soccer player's movements during the game, predicting whether a person is going to cross the street or not, or to track and analyze the movement of microscopic organisms in time-lapse microscopy images, etc. In this article, we will go through a state of the art Offline tracking framework for solving the problem of MOT. The approach that we are about to discuss was published in a paper by the researchers at the Dynamic Vision and Learning Group at TUM. Their proposed algorithm achieved SOTA on MOT15, MOT16, and MOT17 challenges.