Loss function search for person re-identification
作者：
Highlights：
• We provide an analysis of the margin-based softmax loss and conclude four key properties from the three aspects of gradient, feature, and optimization.
• We carefully design a sampling distribution based on the non-independent truncated Gaussian distributions, through which the sampled loss function conforms to the above four properties.
• Our proposed method has reached state-of-the-art performance on the four commonly used datasets.
摘要
•We provide an analysis of the margin-based softmax loss and conclude four key properties from the three aspects of gradient, feature, and optimization.•We carefully design a sampling distribution based on the non-independent truncated Gaussian distributions, through which the sampled loss function conforms to the above four properties.•Our proposed method has reached state-of-the-art performance on the four commonly used datasets.
论文关键词：Person re-identification,Margin-based softmax loss,Loss function search,AutoML
论文评审过程：Received 6 March 2021, Revised 24 October 2021, Accepted 14 November 2021, Available online 17 November 2021, Version of Record 28 February 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2021.108432
