Deep Mutual Learning for Visual Object Tracking

Abstract Existing deep trackers use deep convolutional neural networks to extract powerful features or directly predict the position of the target. For most deep trackers, it is hard to improve their performance by replacing the original backbone with a more powerfully heavyweight network directly. In this paper, we propose a novel mutual-learning-based training methodology for visual object tracking. By re-training the backbone network with this novel methodology, we can improve the tracking performance simply and effectively. We demonstrate this novel training methodology with two mainstream tracking approaches: correlation-filter-based approach and tracking-by-detection-based approach. First, we reformulate a correlation-filter-based tracker as a fully convolutional network and design an end-to-end tracking framework. With this framework, we can enhance the backbone network in a mutual learning way. Second, we integrate our training methodology into a typical tracking-by-detection-based tracker, and then we improve the tracking performance with a simple offline training process. Extensive experiments on the OTB2013, OTB2015, VOT2017 and LaSOT benchmarks demonstrate that the tracking performance can be improved effectively by using the proposed mutual-learning-based training methodology.