Discriminative Context-Aware Correlation Filter Network for Visual Tracking

Abstract

In recent years, discriminative correlation filter (DCF) based trackers using convolutional features have received great attention due to their accuracy in online object tracking. However, the convolutional features of these DCF trackers are mostly extracted from convolutional networks trained for other vision tasks like object detection, which may limit the tracking performance. Moreover, under the challenge of fast motion and motion blur, the tracking performance usually decreases due to the lack of context information. In this paper, we present an end-to-end trainable discriminative context-aware correlation filter network, namely DCACFNet, which integrates context-aware correlation filter (CACF) into the fully-convolutional Siamese network. Firstly, the CACF is modeled as a differentiable layer in the DCACFNet architecture, which can back-propagate the localization error to the convolutional layers. Then, a novel channel attention module is embedded into the DCACFNet architecture to improve the target adaption of the whole network. Finally, this paper proposes a novel high-confidence update strategy to avoid the model corruption under the challenging of occlusion and out-of-view. Extensive experimental evaluations on two tracking benchmarks, OTB-2013 and OTB-2015, demonstrate that the proposed DCACFNet achieves the competitive tracking performance.