SiamMGT: a mask-guided Siamese network tracker with enhanced local feature learning for object tracking

In the field of object tracking, Siamese networks are efficient, but traditional methods rely on global matching—an approach that suffers from background interference and underutilization of local features. We propose SiamMGT, a mask-guided Siamese network tracker based on SiamGAT. The model consists of three core components: a feature extraction network, a mask-guided region module, and a classification-regression subnetwork. Its innovations lie in three key aspects: first, the design of a Self-similar Soft Threshold Processing module to strengthen the local features of templates; second, the construction of an MGRM to optimize information embedding; and third, the use of a KLCC operation to generate accurate masks. Experiments show that the model outperforms the benchmark model and trackers of the same type on mainstream datasets such as UAV123 and performs well in challenging scenarios such as scale changes. In addition, ablation experiments validate the effectiveness of the proposed modules and mechanisms, along with the model’s strong generalization ability. The source code for this project is released on https://github.com/shenjiangduishou/SiamMGT.