Pseudo loss active learning for deep visual tracking
作者：
Highlights：
• A novel pseudo loss describing the spatial context uncertainty of the tracked target and its surroundings is proposed, based on which the most informative samples from the unlabeled pool could be selected.
• An interval threshold and a temporal penalty are adopted to avoid drastic target appearance variation and reduce the redundancy in the selected samples.
• Theoretical analysis has validated the effectiveness of the proposed pseudo loss for sample uncertainty evaluation.
• PLAL has obtained comparable performance (98%–100%) to the model trained on the entire training dataset with only 3% data and the best performance comparing with multiple active learning methods on different tracking benchmarks.
摘要
•A novel pseudo loss describing the spatial context uncertainty of the tracked target and its surroundings is proposed, based on which the most informative samples from the unlabeled pool could be selected.•An interval threshold and a temporal penalty are adopted to avoid drastic target appearance variation and reduce the redundancy in the selected samples.•Theoretical analysis has validated the effectiveness of the proposed pseudo loss for sample uncertainty evaluation.•PLAL has obtained comparable performance (98%–100%) to the model trained on the entire training dataset with only 3% data and the best performance comparing with multiple active learning methods on different tracking benchmarks.
论文关键词：Active learning,Visual tracking,Pseudo loss,Pseudo label
论文评审过程：Received 21 December 2021, Revised 6 April 2022, Accepted 1 May 2022, Available online 4 May 2022, Version of Record 9 June 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.108773
