Efficient template-separable hierarchical transformer tracking for edge computing

In recent years, transformer-based visual tracking models have demonstrated substantial advancements in modeling capabilities. These approaches utilize the global feature representation of vision transformers to enhance information interaction during tracking. However, their high computational demands pose challenges for efficient deployment on resource-constrained platforms, such as mobile devices and robotic systems. To address this issue, we propose a novel model called OneStar, which refers to a template-branch separable vision transformer tracker designed to balance efficiency and accuracy. Unlike existing one-stream trackers that process the template at every frame, the proposed OneStar model performs template inference for feature extraction and information fusion only during the initialization stage, thereby reducing redundant computations in subsequent frames. Additionally, we devise a guided hierarchical architecture conducive to tracking and introduce a tracking token that effectively guides the weight ratios of multi-scale features. Furthermore, we offer a particularly lightweight model variant tailored for low-power edge computing devices. Extensive evaluations demonstrate that the proposed OneStar model surpasses state-of-the-art real-time trackers while achieving impressive speed. For example, the OneStar model achieves 70.0% Average Overlap (AO), a metric that measures tracking accuracy, on the Generic Object Tracking 10k (GOT-10k) dataset and operates four times faster than other high-performance trackers on edge computing devices.