FusionTrack: Multiple Object Tracking with Enhanced Information Utilization

Multi-object tracking (MOT) is one of the significant directions computer vision. Though existing methods can solve simple tasks like pedestrian well, some complex downstream featuring uniform appearance and diverse motion remain difficult. Inspired by DETR, tracking-by-attention (TBA) method uses transformers to accomplish multi-object tasks. However, there are still issues with TBA within paradigm, such as difficulty detecting objects due gradient conflict in shared parameters, insufficient use features distinguish similar objects. We introduce FusionTrack address these issues. It utilizes a joint track-detection decoder score-guided multi-level query fuser enhance usage information between frames. With improvements, achieves 11.1% higher HOTA metric on DanceTrack dataset compared baseline model MOTR.