Dynamic and reliable subtask tracker with general schatten p-norm regularization

• Published in: Pattern recognition Vol. 120; p. 108129
• Main Authors: Fan, Baojie, Cong, Yang, Tian, Jiandong, Tang, Yandong
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2021
• Subjects: Decision-evaluation strategy; Hyper-graph regularization; Reliable multi-subtask tracking; Weighted schatten [formula omitted]-norm; Decision-evaluation strategy; Weighted schatten p-norm; Hyper-graph regularization; Reliable multi-subtask tracking
• ISSN: 0031-3203; 1873-5142
• DOI: 10.1016/j.patcog.2021.108129

Some multi-task trackers adopt an inaccurate shrink strategy to treat different rank components equally. Thus, their flexibility is vulnerable to some tracking challenges. To resolve this problem, we propose a spatial-aware reliable multi-subtask tracker via weighted Schatten p-norm regularization (SLRT-W), which dynamically chooses the suitable and reliable subset of the whole subtasks for tracking. Its major merits not only assign the flexible weights to different subtask rank components depending on their tracking contribution, but also preserve consistent spatial layout structure and correspondence of layered multi-subtask. Specifically, multiple layered subtasks correspond to different target subregions, they are cooperative and complement. A weighted Schatten p-norm is introduced to adaptively shrink different multi-subtask rank components, and emphasize important components as reliable ones. Then, a structured hyper-graph regularized term simultaneously exploits the intrinsic geometry correspondence among multiple layers of subtasks, and spatial layout structure inside each layer. We devise an alternatively generalized iterated shrinkage method to optimize the multi-subtask Schatten p-norm minimization. Finally, a robust decision-evaluation strategy is developed to choose the reliable multi-subtask tracking combination. Encouraging results on some challenging benchmarks demonstrate the proposed tracker performs favorably in robustness and accuracy, against some state-of-the-art trackers.