Cross-Domain Person Re-Identification Based on Normalized IBN-Net

In existing methods, the data distribution between different domains may have large differences, which can lead to performance degradation in the target domain, and the modelling of feature variability between different domains is not sufficient. Aiming at the problem of severe performance degradation during cross-domain migration, this paper proposes a cross-domain person re-identification method based on normalized IBN-Net. The normalized IBN-Net network introduces instance normalization and batch normalization to handle the feature maps at different scales. First, this study employed the normalized IBN-Net network as the backbone of the ResNet50 network. Second, the SimAM attention mechanism is integrated into the backbone network, which is an attention mechanism for inter-modal fusion,that is mainly used in multimodal data processing tasks,and it learns the spatial attention weights in the person images to obtain person information with more discriminative features. Finally, supervised learning was performed using the cross-entropy loss function during source domain training. Meanwhile, it can obtain the details of the source domain samples using the triplet loss function, thereby improving the classification performance. During target domain training, adaptation to the challenges of viewpoint, illumination, background, and feature distribution differences between samples is achieved by learning variations between the source and target domains. In the test stage, comparative experiments were conducted on two large-scale public data sets, Market-1501 and DukeMTMC-reID, achieving Rank-1 accuracies of 86.6% and 79.3%, respectively, with mean average precision mAPs of 68.7% and 62.6%, respectively. The experimental results show that the proposed method performs better in terms of improving the generalization ability of the model.