Auto-selected Knowledge Adapters for Lifelong Person Re-identification

TL;DR

This work tackles Lifelong Person Re-Identification (LReID), a setting where models must continually learn from non-overlapping datasets while preserving past knowledge, a challenge prone to catastrophic forgetting. It introduces AdalReID, a framework that uses incrementally built, parameter-efficient knowledge adapters and a parameter-free auto-selection mechanism to adaptively fuse domain-specific knowledge; a temperature scheduling strategy further enhances cross-adapter interaction, improving generalization to unseen domains. The method leverages a CLIP-based backbone, inserting adapters into the image encoder and training with image-text losses that bridge visual and textual identity representations. Empirical results across multiple lifelong learning orders show AdalReID outperforms state-of-the-art methods by about 5–10% in $mAP$ on both seen and unseen domains, with modest storage overhead (roughly 18 MB per dataset for adapters) and no reliance on old exemplars, highlighting its scalability and practicality for real-world continual ReID systems.

Abstract

Lifelong Person Re-Identification (LReID) extends traditional ReID by requiring systems to continually learn from non-overlapping datasets across different times and locations, adapting to new identities while preserving knowledge of previous ones. Existing approaches, either rehearsal-free or rehearsal-based, still suffer from the problem of catastrophic forgetting since they try to cram diverse knowledge into one fixed model. To overcome this limitation, we introduce a novel framework AdalReID, that adopts knowledge adapters and a parameter-free auto-selection mechanism for lifelong learning. Concretely, we incrementally build distinct adapters to learn domain-specific knowledge at each step, which can effectively learn and preserve knowledge across different datasets. Meanwhile, the proposed auto-selection strategy adaptively calculates the knowledge similarity between the input set and the adapters. On the one hand, the appropriate adapters are selected for the inputs to process ReID, and on the other hand, the knowledge interaction and fusion between adapters are enhanced to improve the generalization ability of the model. Extensive experiments are conducted to demonstrate the superiority of our AdalReID, which significantly outperforms SOTAs by about 10$\sim$20\% mAP on both seen and unseen domains.

Figures (6)

• Figure 1: The illustration of our motivation. (a) Rehearsal-free: with no past knowledge is available, it inevitably leads to the conflict between different knowledge. (b) Rehearsal-based: exemplars will introduce extra storing costs and limited samples for each task may lead to sub-optimization. (c) By contrast, we introduce auto-selected knowledge adapters to learn domain-specific knowledge and then dynamically assign appropriate adapters for LReID. (d) Our approach outperforms previous LReID competitors with a significant advantage on both seen and unseen domains. Please refer to Sec. \ref{['subsec:setup']} for abbreviations of datasets. Best viewed in color and zoomed in.
• Figure 2: Overview of our framework AdalReID. (a) Our proposed Auto-Selected Knowledge Adapter (ASKA) is a parameter-efficient module, which can be plugged into some parameterized layers of each image encoder block. (b) The ASKA module consists of two key components, the knowledge adapters and the knowledge auto-selector. During training, we incrementally build adapters to learn domain-specific knowledge. (c) Our parameter-free auto-selector estimates the knowledge similarity via the statistical distribution, which is incorporated with the temperature scheduling to adaptively select appropriate adapters for the current data or the testing set.
• Figure 3: Choices of temperature scheduling. (a) We visualize several choices of monotonically decreasing function $g(\cdot)$. (b) mAP performance of using different functions ($a=0.5$, $b=0.1$). (c) Given cosinoidal function, we show mAP of varying $a$ and $b$. Best viewed in color and zoomed in.
• Figure 4: Results on seen domains by varying different values of $r$ and $\alpha$ in Eq. \ref{['eq:incremental_lora']}.
• Figure 5: Confusion matrix of knowledge similarity between different ReID datasets.