Deep mutual learning for visual object tracking

Haojie Zhao, Gang Yang, Dong Wang, Huchuan Lu

Abstract: Existing deep trackers use deep convolutional neural networks to extract powerful features or directlypredict the position of the target. For most deep trackers, it is hard to improve their performance byreplacing the original backbone with a more powerfully heavyweight network directly. In this paper, wepropose a novel mutual-learning-based training methodology for visual object tracking. By re-trainingthe backbone network with this novel methodology, we can improve the tracking performance sim-ply and effectively. We demonstrate this novel training methodology with two mainstream tracking ap-proaches: correlation-filter-based approach and tracking-by-detection-based approach. First, we reformu-late a correlation-filter-based tracker as a fully convolutional network and design an end-to-end trackingframework. With this framework, we can enhance the backbone network in a mutual learning way. Sec-ond, we integrate our training methodology into a typical tracking-by-detection-based tracker, and thenwe improve the tracking performance with a simple offline training process. Extensive experiments onthe OTB2013, OTB2015, VOT2017 and LaSOT benchmarks demonstrate that the tracking performance canbe improved effectively by using the proposed mutual-learning-based training methodology.