Re-Identification and Tracklet-Plane Matching for Multi-Object Tracking and Segmentation

As a significant extension to the widely applied multiobject tracking (MOT), multi-object tracking and segmentation (MOTS) introduces segmentation masks as additional information to the bounding box based tracking task, which brings more natural descriptions of the scene and offers a potential solution to the occlusions between objects. Nevertheless, the segmentation information raises new challenges to conventional tracking methods; effective utilization of pixel-wise masks thus becomes a vital topic in MOTS. In this paper, we focus on the tracking task itself with pre-defined detection and segmentation results in the scene, and propose a novel framework that integrates segmentation-based feature extraction, short tracklet construction, and trackletplane matching for long trajectories. To make our method robust for varied object scales, we also propose a depthaware instance filtering strategy, which dynamically adapts the filtering threshold. During the 5th BMTT MOTChallenge Workshop of CVPR 2020, our method achieves the 2nd place on the MOTS20 leaderboard.