Prompt-Based Transformer for Generalizable Person Re-identification with Image Masking

Domain generalizable (DG) person re-identification (ReID) aims to perform well on the unseen target domains by training on multiple source domains with different distribution, which is a realistic but challenging problem. Existing DG person ReID methods have not well explored the domain-specific knowledge based on the Transformer. In this paper, we propose a Prompt-based transformer framework which embeds domain-specific knowledge into different domain prompts, which are optionally optimized by different source domains. Furthermore, we exploit a pretext task of masking and predicting for DG ReID to broaden the understanding of model about data by learning from the signals of the corresponding matching image, which enables interaction between image pairs and improves the ability of generalization. Extensive experiments demonstrate that our method achieves state-of-the-art performances on the popular benchmarks.