This paper addresses the problem of Lifelong Person Re-identification (LReID) in continuous data learning. LReID presents a trade-off between within-domain differentiation (subtle individual differences, e.g., clothing, accessories) and cross-domain differences. Existing methods primarily focus on reducing between-domain differences through knowledge distillation, but tend to overlook within-domain differences. To balance within-domain differentiation and cross-domain differences, this paper proposes a novel domain-consistent representation learning (DCR) model that leverages global and attribute-specific representations. At the within-domain level, we leverage the complementary relationship between global and attribute-specific representations to enhance differentiation between similar identities. To address the forgetting problem caused by excessive within-domain differentiation, we propose attribute-based forgetting prevention (AF) and knowledge consolidation (KC) strategies. Experimental results demonstrate that the proposed DCR model outperforms state-of-the-art LReID methods.
