Pedestrian re-identification based on Swin Transformer
Zifei Qin, Peishun Liu, Yibei Liu, Haiping Duan, Li Fei-Fei, Han Wang · 2022 3rd International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (ICBAIE) · 2022
At present, most mainstream pedestrian re-identification(ReID) methods are based on convolution neural network(CNN). However, CNN has the problems of information loss caused by down-sampling algorithm and remote dependence modeling. Transformer architecture has achieved good performance in various visual tasks recently. In order to explore its ability in ReID task, we designed a ReID model based on Swin Transformer architecture, and improve it by combining multi-scale feature fusion. Specifically, Swin Transformer is used as a backbone network to extract image features, and design two modules to enhance feature perception. (1) Shallow feature extraction module (SFE) is embedded in the initial position of the model to generate robust shallow feature expression. (2) Multi-scale feature fusion module (MFM) extracts signs from different stages of the model, and further enhances the feature extraction ability of the model by fusing different levels of features. Based on experimental results, the proposed model based on Swin Transformer achieves a higher recognition accuracy than the mainstream CNN-based pedestrian re-identification method.