Deformation Samples Generated Network for Robust Visual Tracking

Object tracking has been extensive applied in video surveillance. However, deformation always makes the tracker have a drift problem, how do we learn a tracker classifier that is robust to deformations? We think the performance of existing trackers using deep classification networks is limited by the training data with different challenging factors. Collecting even larger training dataset is the most intuitive paradigm, but it may still can not cover all situations and the positive samples are still monotonous. To address this problem, in this paper, we propose a novel network to generate deformation samples via deformation samples generated network (DSGN). The goal of our tracker is to generate deformation samples which are difficult for the tracker to classify. In our framework both the classifier and DSGN are learned in a join-t manner. Compared with the recent state-of-the-art trackers, our proposed algorithm illustrates outstanding performance relative to existing tracking benchmarks.