Cloth-debiasing with Stable Diffusion in Cloth-changing Person Re-identification

In the current study of cloth-changing person re-identification (CC-ReID), the misidentification rate is significantly high for different individuals wearing similar attire, due to biases in clothing features. The generation model can unify the clothing feature space, minimizing interference from clothing color and type, and enabling model to concentrate on extracting clothing-irrelevant features. However, the current use of Generative Adversarial Networks (GANs) for changing clothes in CC-ReID faces challenges in addressing complex discrepancies between the generated images and the original images, resulting in unstable outcomes when changing the same clothing for pedestrians with varying postures and clothing types. Consequently, we generate cloth-changing pedestrian images with consistent clothing based on a stable diffusion model controlled by body keypoint information, ensuring the images conform to the geometric structure of human bodies. Additionally, we employ threshold filters to refine these images, aiming to construct a high-quality CC-ReID dataset with consistent clothing styles. Meanwhile, we improve the CC-ReID model by introducing centroid loss to increase inter-class differences, thereby maximizing the model’s ability to distinguish between pedestrians wearing similar clothing. Extensive experiments demonstrate that our approach outperforms previous methods, achieving a 6% increase in Rank-1 and a 4.4% increase in mAP on the PRCC dataset compared to the baseline.