Siamese Tracker with Multi-Stage Unified Backbone Network

Backbone network plays an essential role in Siamese-based trackers, which can provide rich representational features for the subsequent target state prediction. Among the current mainstream backbones, convolution-based ones often fail to obtain global long-range dependencies, and Transformer-based ones tend to have high computational complexity. In this paper, a Multi-Stage Unified Backbone is proposed for the feature extraction of Siamese tracker. It organically unifies convolution and Transformer, can extract both local detail information and global semantic information, and run with low memory consumption. Experiments on multiple tracking benchmarks, GOT-10K, LaSOT, OTB-100, demonstrate the advantages of this backbone, which can extract more robust target features and effectively address the challenges of target occlusion, deformation, and fast motion.