Deep features for person re-identification on metric learning
Wanyin Wu,Dapeng Tao,Hao Li,Zhao Yang,Jun Cheng
DOI: https://doi.org/10.1016/j.patcog.2020.107424
IF: 8
2021-02-01
Pattern Recognition
Abstract:<p>Person re-identification, a branch of image retrieval, is an increasingly important public safety application. When monitoring larger areas, it is crucial to correctly match the same person in different camera views. With the emergence of deep learning and large-scale data, metric learning has significantly improved person re-identification performance, but the extent to which deep features affect metric learning performance is unknown. However, given the large number of approaches, datasets, evaluation indices, and experimental environments, comparing metric learning methods directly is difficult. To obtain a more comprehensive empirical evaluation of the person re-identification, here we summarize the different types of features and metric learning approaches from a label attributes perspective. Then, by combining advanced approaches to data enhancement and feature extraction, we conduct comprehensive experiments on metric learning methods with two datasets. For fairness, all methods use a unified code library that includes two data enhancement schemes, eight feature extraction algorithms, and eight metric learning methods. Our results show that, the relations of loss function with deep feature space and metric learning.</p>
computer science, artificial intelligence,engineering, electrical & electronic