Camera-invariance correlation learning and inter-domain-specific distinct representation for person re-identification

Unsupervised domain adaptation (UDA) in person re-identification (ReID) experiences performance degradation due to two primary factors: (1) the information changes across cameras among intra-domain identity samples within the source and target domains; and (2) the inter-domain gap when applied to an unexplored image domain, which makes it a challenging task. The majority of current methodologies predominantly address the inter-domain gap while overlooking the impact of the intra-domain issue. To overcome this limitation, this paper proposes an effective invariance representation ReID (IR-ReID) network that encompasses camera-invariance correlation learning (CICL) and Transformer-based, inter-domain-specific discriminative representation (IDSR), supported by curriculum learning, with the aim of concurrently addressing both of these pivotal issues. Specifically, a CICL module is designed to explore intra-domain distinct features of an individual captured by multiple cameras across the source and target domains. Secondly, an IDSR module is proposed to address the inter-domain gap issue. Thirdly, a curriculum learning-driven strategy is introduced for sample training, systematically guiding our network in learning samples progressively, ranging from easy to difficult instances. Meanwhile, a cluster-based constraint loss is presented to further optimize the clustering results. Finally, extensive experiments are conducted on three well-known datasets, which demonstrate the effectiveness and superiority of our proposed approach in comparison to the prevailing state-of-the-art methodologies.