Multi-Task Driven Feature Models for Thermal Infrared Tracking

• DOI: 10.1609/aaai.v34i07.6828
• Publication Date: 2020-06-29T18:34:42Z
• Authors: Qiao Liu; Xin Li; Zhenyu He; Nana Fan; Di Yuan; Wei Liu; Yongsheng Liang
• Keywords: Discriminative model; BitTorrent tracker; Feature (linguistics); RGB color model

Existing deep Thermal InfraRed (TIR) trackers usually use the feature models of RGB for representation. However, these learned on images are neither effective in representing TIR objects nor taking fine-grained information into consideration. To this end, we develop a multi-task framework to learn TIR-specific discriminative features and correlation tracking. Specifically, first an auxiliary classification network guide generation distinguishing belonging different classes. Second, design aware module capture more subtle same class. These two kinds complement each other recognize levels inter-class intra-class respectively. using matching jointly optimized tracking task. In addition, large-scale training dataset train adapting model domain. Extensive experimental results three benchmarks show that proposed algorithm achieves relative gain 10% over baseline performs favorably against state-of-the-art methods. Codes available at https://github.com/QiaoLiuHit/MMNet.