CATrack: Combining Convolutional and Attentional Methods for Visual Object Tracking

The current popular tracking frameworks prioritize the modeling of global relationships while neglecting research on local feature extraction. This paper introduces CATrack, a novel approach for visual object tracking that integrates convolution and attention into a unified framework. In contrast to prior research, it constructs a tracking module using a unified framework that incorporates convolution and attention as its core components. Our method effectively bridges the gap between the two calculation methods. It improves the ability to extract fundamental features, integrates past experience in the tracking field more effectively, while balancing local and global contextual information. The proposed tracker achieves competitive performance on 5 challenging short-term and long-term benchmarks and can run at real-time speed.