Hybrid Siamese-attention Robust Tracker for SOT
Tarek S. Ghoniemy, Mohamed Mahmoud Fouad · 2022
Single object tracking (SOT) finds the object correspondence in a spatio-temporal domain, and incorporates various applications in Computer Vision. The Siamese trackers have recently improved the accuracy of visual tracking, however, with limited accuracy that would be enhanced using an attention-based model. In this paper, an integrated attention-based Siamese (Att-siamMask) network is presented for SOT to improve the Siamese discrimination ability at both semantic and textural feature processing levels. In the proposed approach, the correlation step is performed to enhance the matching results of the sub-Siamese network. Experimentally, results show the robustness of the Att-siamMask model that outperforms the competing Siamese-based trackers by an average improvement of 1.4%, 19.5% and 27% in terms of accuracy, expected average overlap and robustness metrics, respectively, using the two standard video sequences VOT2016 and VOT2018.