Learning Frequency-Based Disentanglement and Filtering for Generalizable Person Re-identification

Domain Generalization (DG) in Person Re-identification (ReID) tackles the task of testing in unseen domains without using target domain data during training. Existing DG ReID methods achieve impressive performance with unified ensemble models or multi-expert hybrid networks. However, as the number of source domains increases, complex relationships between training samples result in domain-invariant characteristics with spurious correlations, impacting further generalization. To address this, we propose a Bilateral Frequency-Aware Network(BFAN) that leverages spectral feature correlation learning for discriminative hybrid features. BFAN includes a Bilateral Frequency Component-guided Attention (BFCA) module to capture semantic information from diverse frequency features and fuse it with spatial features. Additionally, a Fourier Noise Masquerade Filtering (FNMF) module is introduced to suppress non-generalization-supporting components in the frequency domain. Extensive experiments on various datasets demonstrate our method's notably competitive performance.