Distribution-aware Knowledge Unification and Association for Non-exemplar Lifelong Person Re-identification

TL;DR

This paper tackles lifelong person re-identification by addressing catastrophic forgetting and cross-domain adaptation without storing old samples or using distillation. It introduces DKUA, a distribution-aware framework that propagates domain-specific representations through a domain-style encoder and transfer modules, while dynamically unifying them into a cross-domain center via adaptive knowledge consolidation. Further, unified knowledge association explicit­ly models inter-domain relationships, and distribution-based knowledge transfer aligns current domain distributions with the unified cross-domain distribution. Experiments on seen and unseen domains demonstrate substantial improvements over state-of-the-art rehearsal-based and rehearsal-free methods, with strong generalization and competitive efficiency, supported by ablations and visualizations. The approach offers a privacy-preserving, exemplar-free path to robust lifelong ReID in continuously evolving environments, and code is released at the authors’ GitHub repository.

Abstract

Lifelong person re-identification (LReID) encounters a key challenge: balancing the preservation of old knowledge with adaptation to new information. Existing LReID methods typically employ knowledge distillation to enforce representation alignment. However, these approaches ignore two crucial aspects: specific distribution awareness and cross-domain unified knowledge learning, both of which are essential for addressing this challenge. To overcome these limitations, we propose a novel distribution-aware knowledge unification and association (DKUA) framework where domain-style modeling is performed for each instance to propagate domain-specific representations, enhancing anti-forgetting and generalization capacity. Specifically, we design a distribution-aware model to transfer instance-level representations of the current domain into the domain-specific representations with the different domain styles, preserving learned knowledge without storing old samples. Next, we propose adaptive knowledge consolidation (AKC) to dynamically generate the unified representation as a cross-domain representation center. To further mitigate forgetting, we develop a unified knowledge association (UKA) mechanism, which explores the unified representation as a bridge to explicitly model inter-domain associations, reducing inter-domain gaps. Finally, distribution-based knowledge transfer (DKT) is proposed to prevent the current domain distribution from deviating from the cross-domain distribution center, improving adaptation capacity. Experimental results show our DKUA outperforms the existing methods by 7.6%/5.3% average mAP/R@1 improvement on anti-forgetting and generalization capacity, respectively. Our code is available at https://github.com/LiuShiBen/DKUA.

Figures (6)

• Figure 1: (a) Rehearsal-based approaches enforce representation alignment by storing old samples, raising data privacy concerns. (b) Rehearsal-free methods only model the style of the last domain, limiting the model’s long-term anti-forgetting capacity. (c) Our DKUA learns domain-specific styles generated by a domain-style encoder (DSE) including multiple transfer modules (TM) to achieve knowledge consolidation at both the instance and domain levels, enhancing anti-forgetting and generalization capacity.
• Figure 2: The pipeline of our DKUA framework. At the current training domain, the distribution-aware model transfers the robust representation of each instance to domain-specific representations $\{\theta^i\}_{i=1}^{t}$ with the dedicated domain styles. The AKC aims to learn the unified representation $\theta$ dynamically unify domain-specific representations while ensuring consistency alignment inter-domain. The UKA explores the unified representation as a bridge to explicitly model inter-domain associations $\{A_{u-i}\}_{i=1}^{t}$. The DKT prevents the current domain distribution from deviating from the unified distribution.
• Figure 3: Anti-forgetting curves on Market-1501.
• Figure 4: Generalization tendency on Unseen domain.
• Figure 5: The t-SNE visualization results of the unified representation across five domains.