Figure-Aware Tracking under Occlusion from Monocular Videos

摘要

In this paper, we propose a figure-aware tracking framework incorporating figure/ground repulsive forces in a simultaneous detect let classification and clustering problem in the joint space of detect lets and trajectlets for monocular videos. Without depth/disparity, fine-grained trajectlets tend to cause under-segmentation of similarly moving objects or over-segmentation of articulated objects into rigid parts. Detect lets represented by the bounding boxes only help avoiding under-segmentation of similarly moving objects under canonical pose, while do no good for improving the over-segmentation problem. Pose estimation, though not accurate, is often sufficient to segment human torso from its backgrounds and induce figure/ground repulsions, which could reduce the risk of both under-segmentation and over-segmentation. Figure-aware mediation encodes repulsive segmentation information in trajectory affinities and provides more reliable model aware information for detect let classification. Our algorithm can track objects through sparse, inaccurate detections, persistent partial occlusions, deformations and background clutter.