Attention Modulated Multiple Object Tracking with Motion Enhancement and Dual Correlation

The one-shot multiple object tracking (MOT) framework has drawn more and attention in the MOT research community due to its advantage inference speed. However, accuracy of current approaches could lead an inferior performance compared with their two-stage counterparts. reasons are two-fold: one is that motion information often neglected single-image input. other detection re-identification (ReID) two different tasks focuses. Joining at training stage a suboptimal performance. To alleviate above limitations, we propose network named Motion Correlation-Multiple Object Tracking (MAC-MOT). MAC-MOT introduces enhance module (MEA) dual correlation (DCA). MEA performs differences on adjacent feature maps which enhances motion-related features while suppressing irrelevant information. DCA focuses decoupling task strike balance reduce competition between these tasks. Moreover, symmetry core design idea our proposed reflected Siamese-based deep learning backbone networks, input stream images, as well module. Our approach evaluated popular benchmarks MOT16 MOT17. We demonstrate can achieve better than baseline state arts (SOTAs).