Exploiting spatial relationships for visual tracking

Abstract

Transformer has been widely applied to visual tracking tasks, and the performance of object tracking keeps getting better since the attention mechanism excels in capturing long-range dependencies. However, conventional attention mechanisms can only capture feature dependencies from a single dimension, limiting their ability to fully leverage the spatial position information that is critical for accurate target localization. In this work, we propose a novel axial attention mechanism that is specifically designed for object tracking tasks that require precise target positioning. The axial attention mechanism utilizes both row attention and column attention, enabling it to capture global feature dependencies from both dimensions. Additionally, we propose a low-complexity yet highly effective feature fusion network that employs a customized dual attention mixing strategy. Our CRTrack outperforms all state-of-the-art trackers on GOT-10k, LaSOT, TrackingNet, and UAV123, demonstrating the superiority of the framework. Notably, the MACs and Params of our proposed feature fusion network are only 40% and 24% of those of the Transformer-based feature fusion network, respectively.