The Good, the Bad, and the Motion of Parts for deformable object tracking

Abstract

Visual object tracking has been a popular problem for several decades and has undergone many improvements, but tracking deformable objects is still a challenging problem. One possible approach is to track the parts of an object, and in this paper, we propose a novel deep-learning-based framework for part-based tracking. Our framework consists of three different types of networks, i.e., the part generator (PG), the part evaluator (PE), and the part tracker (PT). Given a bounding box of an object in a frame, PG generates good part candidates of the object and they are individually tracked by PT. Then, PE evaluates each of the tracked parts to decide whether it still represents a part of the object well. Bad parts are discarded based on the results of PE, and PG supplements the pool of parts with more candidates to continue the tracking procedure. These steps are repeated until the end, and the organization of the parts is constantly changed in the process. The experiments show that the proposed method shows the state-of-the-art performance in regard of robust deformable object tracking, while showing competitive performance in conventional tracking scores.