Cross-modality Person Re-Identification Based on Dual Attention Mechanism and Multi-scale Features

In order to solve the huge modal difference between visible and infrared images in cross-modality person re-identification, a cross-modality person re-identification network based on dual attention mechanism and multi-scale features was proposed. Firstly, in order to mine the potential spatial feature representation, an efficient multi-stage weighted feature space aggregation module was proposed to integrate features from different stages. Secondly, based on the concept of combining global features and local features, the network is designed as a double-branch structure, and the multi-scale features are complemented by learning information of coarse-grained and fine-grained levels, so as to significantly enhance the feature expression ability of the network. Finally, by learning the discriminating local aggregation features to enhance the anti-interference ability of the global features under unstable factors such as noise and noise, the intra-modality weighted-part aggregation module is introduced. On the SYSU-MM01 dataset, rank-1 and mAP reach 68.24% and 63.86%, respectively, and on the RegDB dataset, the above two parameters reach 92.08% and 88.84%, respectively. Experimental results show that the proposed method is better than most of the existing advanced methods, which fully proves the effectiveness of the proposed method.