Invariance Learning Under Uncertainty for Single Domain Generalization Person Re-Identification

Generalizing the person re-identification (re-ID) algorithms to unseen domains is an important research topic for intelligent video surveillance. However, during the training of single-source domain generalization (S-DG), some difficulties, such as data limitation, single data style, and label-specific interference, can not only influence the representation learning of the model but also lead to poor generalization performance. This article presents a novel label and distribution uncertainty (LDU) model for single-source domain generalization person re-ID. The primary objective is to improve generalization abilities and acquire more invariance representations in the uncertainty and variable style. Specifically, a lightweight but effective LDU model was established, which consistently applied the concept of label uncertainty (LU) throughout the training process. The true labels and cluster-based pseudo-labels are used to extract id-relevant and generalization information from features that mix with domain-specific information. Moreover, the residual blocks with perturbation BN layer (Res-PBN) based on the distribution uncertainty (DU) are employed to obtain style-robust person representation. The content enhancement (CE) modules are applied after each CNN stage to refine and improve discriminability. By applying these proposed behaviors, the model achieves a better generalization performance on unseen target domains. Extensive experiments on four widely used benchmarks demonstrate that the proposed method achieves competitive performance to state-of-the-art methods.