When correlation filters meet fully-convolutional Siamese networks for distractor-aware tracking

Abstract

Correlation filter and Siamese network based methods are popular in visual tracking due to their competitive performance in both precision and efficiency. However, Siamese network based trackers tend to drift towards similar target regions due to the semantic feature representations and an absence of model update. To overcome this limitation, in this paper, we incorporate correlation filter trackers with the fully-convolutional Siamese network (SiamFC) to achieve distractor-aware tracking. First, the pretrained network in SiamFC is utilized to obtain the response map in a large search area; then a guide line is designed to screen out the potential candidate positions of the target according to the distribution of the computed response map; finally, correlation filter tracker based on HOG features is employed to identify the target location from these candidates. Our method combines the advantage of end-to-end tracking based on deep features in SiamFC and online model-update in correlation filter trackers based on hand-crafted features. Experimental results on OTB and VOT benchmarks show that our proposed algorithm performs favorably in terms of both accuracy and robustness when compared to state-of-the-art trackers while maintaining real-time tracking speed.