Dual Prototype Contrastive learning with Fourier Generalization for Domain Adaptive Person Re-identification

Abstract

Unsupervised domain adaptive (UDA) person re-identification (ReID) focuses on improving the model’s generalization capability from one labeled source domain to the unlabeled target domain. Recently, contrastive learning based on pseudo label assignment has attracted much attention and dominated the field. However, existing methods usually consider bridging the domain gap at the feature level, and they only pull each query instance to get close to its cluster centroid which is stored or computed through a memory bank. The relationship of inter-instance within cluster, especially with the hard instances, is ignored. To this end, we propose a Dual Prototype Contrastive learning with Fourier Generalization (DPCFG) framework for domain adaptive Person Re-identification. First, we introduce the Fourier Generalization (FG) strategy at image level to bridge the domain gap. Concretely, the FG strategy is implemented by replacing the amplitude component of each source domain image with a randomly selected target domain image. Then, the Dual Prototype Contrastive learning (DPC) strategy is further developed to fully exploit the hard positive instances within each cluster. DPC optimizes two contrastive losses by forcing each query to be close to two prototypes: the cluster centroid prototype and the hard positive prototype. The cluster centroid prototype ensures the basic classification accuracy, and the hard positive prototype further improves the classification accuracy by dynamically depicting a certain class boundary for each cluster as the model iterates. Experimental results on the real-world datasets, Market, DukeMTMC-reID, and MSMT17, and synthetic dataset PersonX, demonstrate that DPCFG is effective and achieves state-of-the-art UDA person ReID performance.