Distilling the Past: Information-Dense and Style-Aware Replay for Lifelong Person Re-Identification

Abstract

Lifelong person re-identification (LReID) aims to continuously adapt to new domains while mitigating catastrophic forgetting. While replay-based methods effectively alleviate forgetting, they are constrained by strict memory budgets, leading to limited sample diversity. Conversely, exemplar-free approaches bypass memory constraints entirely but struggle to preserve the fine-grained identity semantics crucial for Re-ID tasks. To resolve this fundamental dilemma, we propose an Information-Dense and Style-Aware Replay framework. Instead of storing a sparse set of raw historical images, we fuse the knowledge of sequential data into the pixel space of a compact replay buffer via multi-stage gradient matching and identity supervision. This condensation process not only maximizes the semantic representativeness of limited memory but also naturally conceals original visual details, inherently preserving data privacy. Furthermore, to combat forgetting induced by cross-domain shifts, we introduce a dual-alignment style replay strategy that adapts both current and fused replay samples, harmonizing feature representations across disparate domains. Extensive experiments on multiple LReID benchmarks demonstrate that our method significantly outperforms existing approaches, achieving improvements of +5.0% and +6.0% in Seen-Avg mAP over current state-of-the-art and traditional replay-based methods, respectively, thereby establishing an efficient and robust new baseline for lifelong learning.

Figures (11)

• Figure 1: Instead of storing sparse raw images, our Information-Dense Replay condenses multiple samples into a single representation via pixel-level optimization, breaking physical memory bottlenecks.
• Figure 2: Performance comparison of different methods. We observe that incorporating data replay consistently improves performance, regardless of whether the method initially relies on it, highlighting its effectiveness in LReID scenarios.
• Figure 3: Overall pipeline of our proposed framework.(a) Information-Dense Data Condensation: During the training phase on domain $D_t$, informative replay samples are generated by aligning their gradients with the original data across multiple parameter stages, guided by the ID loss. (b) Bidirectional Style-Semantic Alignment: To mitigate catastrophic forgetting induced by domain shifts, we introduce a dual-alignment mechanism. (b.1) A style transfer network is pre-trained under the supervision of pixel and perceptual reconstruction losses. (b.2) During joint optimization on domain $D_{t+1}$, style transfer is applied bidirectionally—injecting past styles into current streaming data and current styles into buffered replay samples, ensuring robust representation learning across domains.
• Figure 4: Accuracy tendency on each domain and overall average trend across training stages.
• Figure 5: Hyper-parameter analysis of our framework: (a) Gradient matching loss weight ($\alpha$); (b) Perceptual loss weight ($\beta$) for structural preservation; (c) Style alignment weight ($\gamma$); and (d) Information-dense replay weight ($\lambda$). Average performance over five independent runs on Training Order 1 is reported. Dashed lines indicate the default values adopted in our final model.