Distribution Aligned Semantics Adaption for Lifelong Person Re-Identification

TL;DR

This work tackles lifelong person Re-Identification by moving beyond exemplar-based distillation to a pre-trained-semantics paradigm. It introduces Distribution Aligned Semantics Adaption (DASA), combining domain-specific Batch Normalization alignment with a lightweight Semantics Adaption (SA) module and freezing of convolutional layers to preserve general human semantics learned from large-scale pre-training. The approach stores only lightweight BN and SA components per domain, enabling low storage while enabling effective forward knowledge transfer across domains. Empirical results demonstrate state-of-the-art performance in exemplar-free lifelong settings, with strong generalization to unseen domains and reduced resource consumption, highlighting a practical path for deploying lifelong Re-ID systems.

Abstract

In real-world scenarios, person Re-IDentification (Re-ID) systems need to be adaptable to changes in space and time. Therefore, the adaptation of Re-ID models to new domains while preserving previously acquired knowledge is crucial, known as Lifelong person Re-IDentification (LReID). Advanced LReID methods rely on replaying exemplars from old domains and applying knowledge distillation in logits with old models. However, due to privacy concerns, retaining previous data is inappropriate. Additionally, the fine-grained and open-set characteristics of Re-ID limit the effectiveness of the distillation paradigm for accumulating knowledge. We argue that a Re-ID model trained on diverse and challenging pedestrian images at a large scale can acquire robust and general human semantic knowledge. These semantics can be readily utilized as shared knowledge for lifelong applications. In this paper, we identify the challenges and discrepancies associated with adapting a pre-trained model to each application domain and introduce the Distribution Aligned Semantics Adaption (DASA) framework. It efficiently adjusts Batch Normalization (BN) to mitigate interference from data distribution discrepancy and freezes the pre-trained convolutional layers to preserve shared knowledge. Additionally, we propose the lightweight Semantics Adaption (SA) module, which effectively adapts learned semantics to enhance pedestrian representations. Extensive experiments demonstrate the remarkable superiority of our proposed framework over advanced LReID methods, and it exhibits significantly reduced storage consumption. DASA presents a novel and cost-effective perspective on effectively adapting pre-trained models for LReID. The code is available at https://github.com/QizaoWang/DASA-LReID.

Figures (7)

• Figure 1: (a) Comparison of average incremental performance for different methods using the same pre-trained weights. (b) Conventional knowledge distillation pipeline with exemplars. Our proposed DASA paradigm shows great superiority in LReID.
• Figure 2: The framework of DASA. The acquired robust and general human semantics from pre-training are used as shared knowledge, which is kept in the frozen Conv layers. At each training step, we adapt the acquired knowledge from the pre-training to application domain by tuning BN layers and adopting the lightweight Semantics Adaption (SA) modules. During the lifelong evolution process, the previously learned BN and SA are used for initialization in the upcoming domain for forward knowledge transfer, while the old classifier can be discarded without increasing storage burden.
• Figure 3: Comparison of average accuracies at different training steps of Order 1 w.r.t. (a) $\overline{s}_{\rm mAP}$ and (b) $\overline{s}_{\rm R-1}$.
• Figure 4: Comparison of storage consumption for LReID methods. Results are calculated after the last training step.
• Figure 5: Ablation Study of the kernel size for the SA module. The average accuracies of all datasets at the last training step are reported.