Boosting the Generalization Ability of Person Re-identification via Architecture-Level Contrast and Data Augmentation Strategies

To address the challenges of accuracy and robustness in person re-identification (ReID) under complex cross-camera environments, this paper first compares two representative models: the lightweight Light-ReID and the part-based PCB. Through evaluations on three widely used datasets, Market-1501, DukeMTMC-ReID, and MSMT17, Light-ReID consistently demonstrates superior performance, particularly in handling occlusion, background clutter, and variations in viewpoint. Based on this finding, we further explore the impact of a simple yet effective data augmentation strategy (random horizontal flipping) on Light-ReID’s performance. Applying this technique during training makes the model more invariant to orientation changes and achieves additional performance gains. The enhanced version, Light-ReID + Flip, not only significantly outperforms the PCB baseline in both mean Average Precision (mAP) and Rank-1 accuracy but also achieves noticeable improvements compared to the original Light-ReID. These results highlight the importance of integrating lightweight network designs with targeted data augmentation to boost model generalization. This approach provides a practical and efficient solution for improving ReID performance in real-world applications.