Adaptive Style Normalization and Similarity Grouping for Unsupervised Cross-Domain Person Re-Identification

Abstract In unsupervised cross-domain person re-identification, feature extraction often includes identity-irrelevant information, reducing model generalization. Additionally, ignoring potential similarities within the target domain leads to noisy pseudo-labels. To address these issues, this paper introduces a method for cross-domain person re-identification that combines adaptive style normalization with similarity grouping. First, adaptive style normalization is applied on the backbone network to mitigate loss of information caused by instance normalization, thus improving the generalization ability and identity recognition accuracy of the model. Second, similarity grouping is employed to uncover potential similarities in target domain, overcoming the problem of loud pseudo-labels and further improve model performance. Using this method, mAP accuracy increased by 2.27% on the Market-1501 dataset and by 4.26% on the DukeMTMC-ReID dataset, enhancing its application in the real world.