Learning nested attentional feature fusion network for high performance visual tracking

Abstract

Siamese network-based visual tracking has made significant progress in recent years, with correlation calculations playing a central role in these models. However, the inherently linear and localized nature of correlation often leads to substantial semantic information loss and convergence to local optima, thereby limiting the potential for further performance improvements. To address these challenges, we propose a feature fusion network inspired by the Transformer architecture, incorporating nested attention mechanisms to enhance tracking accuracy and robustness. Unlike standard Transformer-based models, our approach refines correlation accuracy by emphasizing correct matches while attenuating incorrect ones through nested attentional representation learning. This enables more effective feature aggregation and information propagation. Our feature fusion network consists of four interdependent modules: ego-context augmentation, short-term feature augmentation, long-term feature augmentation, and cross-feature augmentation. These modules collaboratively fuse features from target templates and search regions, producing semantically rich feature maps superior to those generated by traditional correlation methods. Built on this framework, our proposed model, AiATransT, achieves state-of-the-art performance on five benchmark datasets, validated by extensive experimental evaluations.