Decoupling features via global and local contexts in multi-object tracking

Abstract

Multi-object tracking (MOT) system usually consists of two tasks, object detection and re-identification (ReID). Current MOT methods tend to join detection and ReID in a single network to enhance inference speed. Such one-shot models allow joint optimization of detection and Re-ID via a shared backbone, reducing computation cost. However, the different demands of features between the two tasks in one-shot systems lead to competition in the optimization procedure. The detection task needs the features of the instances with the same class to be similar, while the ReID task needs the features of different instances to be distinguishable. Existing methods address the contradiction by disentangling the features into detection-specific and ReID-specific features. But these methods neglect the discussion of semantic interpretation of disentangling modules. In this paper, we propose a feature decoupling module, Global and Local Context-based Decoupling Module (GLCD), to disentangle features extracted by the backbone into two task-specific features. By extracting global and local contexts, the two tasks can choose different contexts by learnable parameters to enforce each self. We conduct our decoupling module into SOTA one-shot MOT method and experiments show performance improvement.