Visual Tracking With Attentional Convolutional Siamese Networks

Recently Siamese trackers have drawn great attention due to their considerable accuracy and speed. To further improve the discriminability of Siamese networks for visual tracking, some deeper networks, such as VGG and ResNet, are exploited as backbone. However, high-level semantic information reduces the location discrimination. In this paper, we propose a novel Attentional Convolutional Siamese Networks for visual tracking (ACST), to improve the classical AlexNet by fusing spatial and channel attentions during feature learning. Moreover, a response-based weighted sampling strategy during training is proposed to strengthen the discrimination power to distinguish two objects with the similar attributes. With the efficiency of cross-correlation operator, our tracker can be trained end-to-end while running in real-time at inference phase. We validate our tracker through extensive experiments on OTB2013 and OTB2015, and results show that the proposed tracker obtains great improvements over the other Siamese trackers.