RelationTrack: Relation-aware Multiple Object Tracking with Decoupled Representation

Existing online multiple object tracking (MOT) algorithms often consist of two subtasks, detection and re-identification (ReID). In order to enhance the inference speed reduce complexity, current methods commonly integrate these double subtasks into a unified framework. Nevertheless, ReID demand diverse features. This issue results in an optimization contradiction during training procedure. With target alleviating this contradiction, we devise module named Global Context Disentangling (GCD) that decouples learned representation detection-specific ReID-specific embeddings. As such, provides implicit manner balance different requirements subtasks. Moreover, observe preceding MOT typically leverage local information associate detected targets neglect consider global semantic relation. To resolve limitation, develop module, referred as Guided Transformer Encoder (GTE), by combining powerful reasoning ability encoder deformable attention. Unlike previous works, GTE avoids analyzing all pixels only attends capture relation between query nodes few self-adaptively selected key samples. Therefore, it is computationally efficient. Extensive experiments have been conducted on MOT16, MOT17 MOT20 benchmarks demonstrate superiority proposed framework, namely RelationTrack. The experimental indicate RelationTrack has surpassed significantly established new state-of-the-art performance, e.g., IDF1 70.5% MOTA 67.2% MOT20.