Cross Intra-Identity Instance Transformer for Generalizable Person Re-Identification

Person re-identification (ReID) has made significant progress in recent years, but Domain Generalization person re-identification (DG-ReID) remains a challenging task. Current DG-ReID methods update feature extractors and classifier parameters based on single-instance features. Consequently, these methods might not learn the common features between different instances of each identity, thus limiting its generalizability in the unseen target domain. To address this issue, we propose the Cross Intra-Identity Instance (CI3) learning framework based on Transformer. The CI3 framework leverages the mixed cross-attention (MCA) module and dynamic memory bank to extract random intra-identity instances during each iteration to learn additional information beyond the visual features between them. Furthermore, we introduce a task separation layer to eliminate the detrimental information in the backbone network. To bridge the domain gap between the source and target domains, we designed an online generalization feature constraint (OFC) strategy that utilizes constructed pseudo-label information to learn general cross-domain features. Our method effectively enhanced the performance of the DG-ReID model, achieving competitive results in extensive experiments on four benchmark datasets.