DSGA: Distractor-Suppressing Graph Attention for Multi-object Tracking

Guangjie Gao, Yan Gao, Liyang Xu, Huibin Tan, Yuhua Tang

Abstract: Multiple object tracking (MOT) methods based on single object tracking are of great interest because of their ability to balance efficiency and performance on the strength of the localization capability of single-target tracking. However, most of the single object tracking methods only distinguish foreground and background. They are susceptible to the influence of similar interfering objects during localization, while in multiple object tracking scenarios, there are more interfering objects and the influence is more severe. Therefore, we propose a Distractor-Suppressing Graph Attention (DSGA) to learn more discriminative attention by reducing the influence of distractors on learning attention weight features. Furthermore, DSGA is embedded into the basic MOT framework “SiamMOT” formed as DSGA-SiamMOT and applied to multiple object tracking to verify its effectiveness. We conduct experiments on the MOT Challenge benchmark with "public detection", and obtain MOTA 66.65%, IDF1 62.2% accuracy on the MOT17 dataset with 14fps.