Adaptive Feature Selection Siamese Networks for Visual Tracking

Recently, template based discriminative trackers, especially Siamese network based trackers have shown great potential in terms of balanced accuracy and tracking speed. However, it is still difficult for Siamese models to adapt the target variations from offline learning. In this paper, we introduced an Adaptive Feature Selection Siamese (AFS-Siam) network to learn the most discriminative feature information for better tracking. Features from different layers contain complementary information for discrimination. Proposed adaptive feature selection module selects the most useful feature information from different convolutional layers while suppresses the irrelevant ones. Proposed tracking algorithm not only alleviates the over-fitting problem but also increases the discriminative ability. The proposed tracking framework is trained end-to-end. And extensive experimental results over OTB50, OTB100, TC-128, and VOT2017 demonstrate that our tracking algorithm exhibits favorable performance compared to other state-of-the-art methods.