GCNNMatch: Graph Convolutional Neural Networks for Multi-Object Tracking via Sinkhorn Normalization

This paper proposes a novel method for online Multi-Object Tracking (MOT)\nusing Graph Convolutional Neural Network (GCNN) based feature extraction and\nend-to-end feature matching for object association. The Graph based approach\nincorporates both appearance and geometry of objects at past frames as well as\nthe current frame into the task of feature learning. This new paradigm enables\nthe network to leverage the "context" information of the geometry of objects\nand allows us to model the interactions among the features of multiple objects.\nAnother central innovation of our proposed framework is the use of the Sinkhorn\nalgorithm for end-to-end learning of the associations among objects during\nmodel training. The network is trained to predict object associations by taking\ninto account constraints specific to the MOT task. Experimental results\ndemonstrate the efficacy of the proposed approach in achieving top performance\non the MOT '15, '16, '17 and '20 Challenges among state-of-the-art online\napproaches. The code is available at https://github.com/IPapakis/GCNNMatch.\n