@article{Wu2025, 
author = {Huanyu Wu and Yingpin Chen and Changhui Wu and Ronghuan Zhang and Kaiwei Chen},
title = {A multi-scale cyclic-shift window Transformer object tracker based on fast Fourier transform},
year = {2025},
journal = {Electronic Research Archive},
volume = {33},
number = {6},
pages = {3638-3672},
keywords = {object tracking, window attention, cyclic-shift, fast Fourier transform, Transformer tracker},
url = {https://www.sciopen.com/article/10.3934/era.2025162},
doi = {10.3934/era.2025162},
abstract = {In recent years, Transformer-based object trackers have demonstrated exceptional performance in object tracking. However, traditional methods often employ single-scale pixel-level attention mechanisms to compute the correlation between templates and search regions, disrupting object's integrity and positional information. To address these issues, we introduce a cyclic-shift mechanism to expand the diversity of sample positions and replace the traditional single-scale pixel-level attention mechanism with a multi-scale window-level attention mechanism. This approach not only preserves the object's integrity but also enriches the diversity of samples. Nevertheless, the introduced cyclic-shift operation heavily burdens storage and computation. To this end, we treat the attention computation of shifted and static windows in the spatial domain as convolution. By leveraging the convolution theorem, we transform the attention computation of cyclic shift samples from the spatial domain to element-wise multiplication in the frequency domain. This approach enhances computational efficiency and reduces data storage requirements. We conducted extensive experiments on the proposed module. The results demonstrate that the proposed module outperforms multiple existing tracking algorithms regarding performance. Moreover, ablation studies show that the method effectively reduces the storage and computational burden without compromising performance.}
}