Multifactor-Guided Collaborative Learning for Cloth-Changing Person Reidentification

A novel cloth-changing person re-identification (ReID) method utilizes multi-factor to extract cloth-irrelevant identity features. It starts by creating black-cloth images by masking upper clothes and pants from RGB images. Instead of directly learning from these, a teacher ReID model pretrains on them. The original pedestrian images are then used to extract cloth-irrelevant identity features, guided by the pretrained model. The head patch is separately extracted to preserve fine-grained head features. This three-branch framework (black-cloth, original, head) employs an identical network architecture without weight sharing, enhancing the robustness of cloth-irrelevant identity features. The method employs the improved vision transformer (imViT) backbone and achieves relatively good performance on PRCC and VC-Clothes datasets.