Cross-Domain Person Re-Identification Using Value Distribution Alignment

Person re-identification is a technique used to identify individuals across different camera views, with applications in security, surveillance, and public safety. However, it is a challenging task due to factors like varying lighting conditions, camera viewpoints, and the appearance changes of individuals. Consequently, performance degradation often occurs in cross-domain person re-identification when training and testing data come from different datasets. To tackle this issue, we propose a value alignment module that mitigates the impact of brightness differences between domains. The proposed module is simple but effective such that it adjusts the input data at the front end of the model by aligning the brightness distribution of image data with a specific distribution. The performance of the proposed module is evaluated on the cross-domain person re-identification setting using different datasets in training and testing, specifically the Market1501 and CUHK03 datasets. Experimental results demonstrate that applying the proposed module significantly improves the generalization performance of cross-domain person re-identification.