Asymmetric network pseudo labels mutual refinement for unsupervised domain adaptation person re-identification

In the task of unsupervised domain adaptation person re-identification, the traditional symmetric dual-branch network only generates one single feature, which ignores the difference and complementarity of the network and is prone to produce wrong cluster results. To solve this problem, the asymmetric network pseudo labels mutual refinement (ANPR) algorithm is proposed to design an asymmetric network (ASNet), containing diverse features for mutual supervision. One branch of ASNet employs convolutional operations with constrained receptive fields to extract local information, and the other branch uses the contextual transformer block to capture global features. Secondly, this paper constructs the pseudo labels mutual refinement (PMR) module, which generates two sets of clustering results using global and local potential features of unlabeled samples. PMR suppresses the inconsistent clustering results and retains the consistent ones, which gradually improves the quality of pseudo labels for dual-brach mutual supervision. In addition, the feature space loss (FSL) function is designed to focus on the relative distance of samples in the feature layer without relying on the pseudo labels in the task layer, which effectively avoids the influence of noisy pseudo labels in the optimization process. The proposed method is evaluated on three popular datasets, and extensive experimental results demonstrate its effectiveness.