TrackFormer++: End-to-End Multi-Object Tracking with Prior Content Information

Multiple-Object Tracking (MOT) aims to detect objects of interest in a video sequence and maintain their identities across frames. Howerver, it remains a challenging problem in the real-world complex scenes where precise detection and robust data association are required. In this paper, we introduce an end-to-end method with prior content information, which is named TrackFormer++. The method introduces a novel prior content information generation strategy and incorporates a self-attention pooling network to generate a query which provides information about the objects to be detected. With the prior information, it enables precise object localization and accelerates the training convergence. Experimental results on the public dataset MOT17 demonstrate that TrackFormer++ outperforms the benchmark methods in terms of multiple mainstream metrics and remains competitive with current state-of-the-art methods.