PMD-Transformer: A Domain Generalization Approach for Person Re-Identification

Domain generalization person re-identification (DG-ReID) aims to address the performance degradation caused by domain shift between the training domain and unseen target domains. Transformers often demonstrate better generalization ability than CNNs due to their self-attention mechanism and capacity for capturing global features. However, the supervised learning strategy on the source domain makes the Transformer-based ReID model inevitably overfit the bias in some regions. Therefore, a transformer model (PMD Transformer) for DG-ReID is proposed. First, by introducing relative position coding, it prevents the person image block from losing its relative position information when passing through the attention mechanism and improves the extraction of person features. Second, in order to solve the excessive attention of the model to the features of some regions and thus the overfitting phenomenon, the part mask attention (PMA) module is designed to limit the scope of the attention mechanism and improve the computational efficiency. Finally, a new feedforward network (DRFNN) is used to enhance the mastery of spatial information and further improve the generalization performance of the model. The experimental results show that, in the setting where Market is the source domain and Duke is the target domain, the average precision (mAP) and Rank-1 metrics increased by 12.5% and 9.2%, respectively, compared to the optimal algorithm. In the setting where Duke is the source domain and CUHK03-NP is the target domain, the mAP and Rank-1 metrics increased by 10.5% and 8.6%, respectively. The experiments demonstrate that the proposed model exhibits outstanding generalization ability in DG-ReID.