Loading [a11y]/accessibility-menu.js
Learning Attention Through Hierarchical Architecture for Visual Object Tracking | IEEE Journals & Magazine | IEEE Xplore

Learning Attention Through Hierarchical Architecture for Visual Object Tracking


Abstract:

Various attention mechanisms have been extensively adopted in visual object tracking networks. However, the research of Siamese networks with convolutional attention mech...Show More

Abstract:

Various attention mechanisms have been extensively adopted in visual object tracking networks. However, the research of Siamese networks with convolutional attention mechanisms as the main architecture remains limited. In this letter, we propose a novel Siamese tracker that leverages convolutional attention and a hierarchical architecture. By utilizing the deformable convolution, we expand the receptive field to capture more contextual information. In addition, our hierarchical architecture optimizes the attention by hierarchically adjusting the focus of attention based on the significance of features at different levels. On the other hand, we decompose the step of feature compression during the feature-to-result transformation process and integrate it in segments with the feature fusion process to address the information loss and underfitting issues. Extensive experimental results demonstrate that our tracker achieves state-of-the-art performance on multiple benchmarks including LaSOT, TrackingNet, UAV123, and NFS while maintaining a high processing speed of 40 FPS.
Published in: IEEE Signal Processing Letters ( Volume: 31)
Page(s): 186 - 190
Date of Publication: 18 October 2023

ISSN Information:

Funding Agency:


References

References is not available for this document.