Multi Loss Function for Cross-Modality Person Re-Identification

Abstract

For cross-modality person re-identiflcation, the intra-class difference between visible images and infrared images of the same identity is large, and how to reduce this intra-class difference has become the key of cross-modality person re-identification. Therefore, we proposed a multi-loss function for cross-modality person re-identification. Firstly, the global attention mechanism was embedded in the Resnet50 network to retain non-local feature information. Secondly, generalized-mean pooling is used to increase feature information extraction for different fine-grained regions by adjusting parameters. Finally, we design a new total loss function to supervise network learning and improve model accuracy. The proposed method achieves an average accuracy of 54.18% and 78.40% in the SYSU-MM01 and RegDB datasets. The experimental results show that the proposed method can effectively improve the accuracy of cross-modality person re-identiflcation.