Prompt Based Lifelong Person Re-identification

In the real world, training data for person re-identification (ReID) comes in streams and the domain distribution may be inconsistent, which requires the model to incrementally learn new knowledge without forgetting the old knowledge. The problem is known as lifelong person re-identification (LReID). Previous work has focused more on the acquisition of task-irrelevant knowledge and neglected the auxiliary role of task-relevant information in alleviating catastrophic forgetting. To alleviating forgetting and improving the generalization ability, we introduced the prompt to learn task-relevant information, which can guide the model to perform task conditionally. We also proposed a special distillation module for the specific vision transformer structure, which further mitigated catastrophic forgetting. Extensive experiments on twelve person re-identification datasets outperforms other state-of-the-art competitors by a margin of 4.7% average mAP in anti-forgetting evaluation and 7.1% average mAP in generalising evaluation.