PolyTracker: Progressive Contour Regression for Multiple Object Tracking and Segmentation

Sanjing Shen, Hao Feng, Wengang Zhou, Houqiang Li

Abstract: State-of-the-art multiple object tracking and segmentation methods predict a pixel-wise segmentation mask for each detected instance object. Such methods are sensitive to the inaccurate detection and suffer from heavy computational overhead. Besides, when associating pixel-wise masks, additional optical flow networks are required to assist in mask propagation. To relieve these three issues, we present PolyTracker, which adopts object contour, in the form of a vertex sequence along with the object silhouette, as an alternative representation of the pixel-wise segmentation mask. In the PolyTracker, we design an effective contour deformation module based on an iterative and progressive mechanism, which is robust to the inaccurate detection and has low model complexity. Furthermore, benefiting from the powerful contour deformation module, we design a novel data association method, which achieves effective contour propagation by fully mining contextual cues from contours. Since data association relies heavily on pedestrian appearance representation, we design a Reliable Pedestrian Information Aggregation (RPIA) module to fully exploit the discriminative re-identification feature. Extensive experiments demonstrate that our PolyTracker sets the promising records on the MOTS20 benchmark.