AugReID: Transformer-Based Augmentation Person Re-Identification

Person re-identification (ReID) is a crucial component of computer vision that has garnered increased attention in recent years, primarily due to its importance in applications such as video surveillance. However, existing research predominantly focuses on ReID performance in ideal dataset environments, often overlooking the challenges associated with implementing ReID algorithms in real-world deployment scenarios. In these real-world situations, variations in light intensity and resolution between edge-end cameras can considerably hinder the accurate interpretation of person characteristics. To address these challenges, we propose an augmentation strategy to simulate interference in real-world deployment scenarios. Furthermore, we introduce a noise fusion model structure, AugReID, designed to account for the impact of illumination, resolution, and other factors on the ReID task. Extensive experiments on holistic ReID datasets demonstrate our model’s effectiveness. Additionally, our approach surpasses state-of-the-art methods on the Occluded ReID task, which contains more noise-related elements and closely resembles real deployment environments.