Object Tracking Performance Evaluation Method Based on Adaptive Threshold and Background Suppression
Yueping Huang, Shengxiu Zhang, Xiaofeng Li, Ruitao Lu, Qiao Sun · 2022 IEEE 10th Joint International Information Technology and Artificial Intelligence Conference (ITAIC) · 2022
The large difference in target size and the deviation and ambiguity of axis-aligned annotation boxes in public object tracking benchmarks seriously affect the credibility of evaluation indexes based on center location error and overlap rate. To improve the adaptability and credibility of the above evaluation indicators, a target tracking performance evaluation method based on adaptive threshold and background suppression is proposed. The annotated target size is used to adaptively determine the center location error threshold to maintain the consistency of the accuracy evaluation criteria; the background-suppression overlap rate is used to reduce the adverse effect of background information in the axis-aligned annotation boxes on the reliability of performance evaluation indicators. Finally, the effectiveness of the proposed evaluation method is verified by sequences from multiple benchmarks, such as OTB100, TC128, UAV123, and LaSOT.