Transformer Tracking Based on Entropy-Guided Training and Multi-Scale Self-Gating Module

Transformer-based tracking has emerged as a significant method in visual object tracking tasks. However, most research has considerably improved the network architecture to optimize the applicability of Transformers for tracking tasks, resulting in an ever more complex tracking framework. Meanwhile, Transformers struggle to extract meaningful information from low-quality samples, which limits the use of such trackers in real-world situations. This work proposes a Transformer tracker (EMSTrack) using entropy-guided training and a multi-scale self-gating module to resolve the issues above. This research proposes an entropy-guided training strategy for mining hard samples during the model's training phase, enhancing tracking performance without augmenting model complexity, in contrast to prior works that focus on network structure. Initially, an importance weight is assigned to each sample based on the information entropy during the training phase, with the significance of each sample being dynamically modified throughout the training process. Samples with lower information content are assigned greater weights, enabling the model to concentrate on these lower-quality, challenging-toevaluate samples. Additionally, to further augment the feature representation capacity, this research presents a multi-scale selfgating module. Self-gating modules at various sizes enhance the spatial semantic information of the features, while linear procedures consolidate the channel information. This work compares the EMSTrack tracker with several state-of-the-art algorithms on LaSOT, GOT-10k, and TrackingNet, demonstrating that the proposed method achieves superior tracking performance.