Domain generalization person re-identification via style adaptation learning

Domain generalization person re-identification (DG Re-ID) aims to deploy a Re-ID model trained on multiple source domains to unseen domains without adaptation, which is a practical and challenging problem. Due to the significant drop in performance of DG Re-ID methods on unseen target domains, Mixture of Experts (MoE) is used to set up an expert branch for each domain to learn its unique features and thus improve the identification performance on unseen target domains. However, most of the MoE-based DG Re-ID methods only aggregate multiple branch features to simulate the unseen target domain and ignore the role of style in classification recognition among different domains. To address these problems, we propose Style-Adaptive Learning DG Re-ID (SALDG) method, which includes Global Branch, Expert Branches, Class-level Style-adaptive Learning (CSL) module, and Domain-level Style-adaptive Learning (DSL) module. The Global Branch and the Expert Branches extract the global feature and different domain style features respectively. The CSL module can learn the style information of each domain, and the DSL module adaptively mixes the style information of each domain learned by the CSL module to form more complex style information and enhance style diversity in the global feature space. Through extensive experiments, compared with the state-of-the-art methods, the mAP and R-1 accuracy of SALDG on Market1501, DukeMTMC, CUHK03, MSMT17 datasets are improved by approximately 4.1% and 5% respectively.