Wavelet-domain feature decoupling for weakly supervised multi-object tracking

We present a wavelet-domain feature-decoupling Transformer-based tracking network for the weakly supervised MOT task (FDMOT). Our FDMOT has two improvements over the previous weakly supervised methods. First, FDMOT decouples noisy intermediate features caused by noisy pseudo identity labels in the wavelet domain, extracting discriminative features for accurately detecting and identifying multiple targets. Second, FDMOT further improves the noise-decoupled embedding features into the well-refined ones with the cooperation of the three feature-decoupling Transformer-based branches, which can accurately identify and track heavily occluded targets in crowded scenes. Experimental results show the superiority of FDMOT compared with several state-of-the-art supervised and weakly supervised MOT methods.