Descriptive Textual Prompted Prototype Compensation for Lifelong Person Re-identification

Lifelong person re-identification (LReID) aims to continuously learn new identities from evolving data streams in real-world settings, while avoiding the loss of previously learned knowledge. The major challenge in LReID is the misalignment of the feature space caused by regular model updates. To address this, existing methods typically retain a small set of historical exemplars or leverage knowledge distillation to reduce feature drift. However, these approaches risk undermining the model's capacity to learn new knowledge effectively. To overcome these challenges, we propose a novel framework called Descriptive Textual Prompted Prototype Compensation. This framework alleviates feature drift, mitigates catastrophic forgetting, and enables concurrent learning of new tasks. Specifically, we introduce a Prompted Prototype Compensation (PPC) module that achieves fine-grained semantic drift compensation by leveraging multimodal visual and textual prototypes generated by CLIP. Additionally, we implement a Boundary-aware Prototype Correction (BPC) module that dynamically samples compensated pseudo-features from old identities that are at risk of being confused with others. This helps maintain clear decision boundaries, enhancing the model's adaptability. Through extensive experiments across multiple benchmarks, our approach significantly outperforms existing methods, demonstrating its superior effectiveness and capability to handle real-world challenges in LReID.