Learning diverse fine-grained features for thermal infrared tracking

Existing feature models used in thermal infrared (TIR) tracking struggle to get strong discriminative features of TIR objects, because TIR image has few details and low contrast. This characteristic makes the existing TIR tracking methods easy to drift to similar distractors. To tackle this problem, we introduce a novel diverse fine-grained feature network for TIR tracking. Our proposed method emphasizes extracting fine-grained features from multiple local regions of the target to improve its ability to discriminate against distractors. Specifically, our feature model consists of a specific-designed fine-grained feature network architecture and an auxiliary diversity loss function. Firstly, the fine-grained feature network explores the subtle clues of the infrared target through a mask suppression mechanism. This mechanism can force the network to learn subtle cues of the target. Secondly, to guarantee the learned fine-grained features are various, we propose a diversity loss to force all fine-grained features to be unique. These two modules help the feature model learn diverse fine-grained features from two complementary aspects. To verify the effectiveness of the two modules, we evaluate them on four benchmarks. The relevant experimental results prove that our proposed method achieves the best performance compared to the state-of-the-art methods.