Learning Discriminative Representations through an Attention Mechanism for Image-based Person Re-identification

Jing Liu1, 2 and Guoqing Zhou2

  1. School of Computer Science, Weinan Normal University,
    Weinan, 714099, Shaanxi, China.
    liujing8318@mail.nwpu.edu.cn
  2. School of Computer Science, Northwestern Polytechnical
    University, Xi’an, 710072, Shaanxi, China.
    Zhouguoqing@nwpu.edu.cn

Abstract

Over the past years, person re-identification has been obtaining various attentions in computer vision tasks. However, existing methods mainly focus on building massive number of deep architecture layers, which is unsuitable for extracting the robust features for person re-ID. In this paper, we present a novel hybrid framework PGAN, through which the discriminative representations can be learned for person re-ID. Specifically, a novel self-attention method named channel-wise attention mechanism is adopted to learn the informative representations from the patch-network and global network, respectively. In addition, CSwin Transformer is exploited to re-extract the discriminative features from the residual blocks. We obtain a mAP of 81.8% and 80.3% of the labeled and detected dataset on the CUHK0-NP dataset. And we obtain a mAP of 83.4% and 91.3% on the DukeMTMC and Market-1501 datasets respectively. Comprehensive experiments are performed on the three datasets, (Market-1501, DukeMTMC-reID and CUHK03-NP), demonstrating the efficiency of the introduced approach.

Key words

person re-identification, channel-wise attention, deep learning

Digital Object Identifier (DOI)

https://doi.org/10.2298/CSIS230829044L

Publication information

Volume 21, Issue 4 (September 2024)
Year of Publication: 2024
ISSN: 2406-1018 (Online)
Publisher: ComSIS Consortium

Full text

DownloadAvailable in PDF
Portable Document Format

How to cite

Liu, J., Zhou, G.: Learning Discriminative Representations through an Attention Mechanism for Image-based Person Re-identification. Computer Science and Information Systems, Vol. 21, No. 4, 1483–1498. (2024), https://doi.org/10.2298/CSIS230829044L