MSNet: a lightweight multi-scale deep learning network for pedestrian re-identification

作者：Pan, Keyu; Zhao, Yishi; Wang, Tao; Yao, Shihong^*

来源：Signal Image and Video Processing, 2023, 17(6): 3091-3098.

DOI：10.1007/s11760-023-02530-1

• Pedestrian re-identification
• Multi-scale feature
• Lightweight
• Multi-branch
• Segmentation

Pedestrian re-identification is highly dependent on discriminative features that enable images to encapsulate an arbitrary combination of multiple scales by different spatial scales. However, current models divide the scale by mechanical horizontal segmentation, which inevitably degenerate the re-identification performance. In this paper, we propose a novel multi-scale network (MSNet) to extract a certain scale feature map through different branches before segmentation. The branches utilize backbone networks composed of multi-scale residual blocks to extract features at different scales. Moreover, the specific segmentation method of the feature map is also based on its scale, which is opposite to the method of the first segmentation and then determine the scale. Moreover, MSNet significantly shortens the training and testing time owing to its lightweight design. Experimental results evidently demonstrate that the proposed MSNet shows superior performance in terms of accuracy, efficiency, and robustness on three open-source data sets, compared with other models. Codes are available at https://github.com/PKY-IMO/MSNet.