Intermittent Learning Strategy via Self-supervised Learning for Sketch-visible Person Re-identification

This paper proposed an ILS (Intermittent Learning Strategy) for sketch-visible person re-identification.The ILS method performs intermittent epochs between self-supervised learning and person re-identification.It can be used as regularization for the training process and reduce the overfitting problem on person re-identification problems by disrupting the learning process for more simple tasks.We use a simple vertical jigsaw pre-task for self-supervised learning by dividing the person's image into several vertically cropped regions.The cropped regions were then labeled and used for self-supervised learning.We perform the experiments using two deep learning architectures: Swin Transformer and Swin Transformer V2 classifiers, along with the PKU-Sketch-ReID and Market-Sketch datasets.Experiments on two different sketch-visible person re-identification datasets show that the proposed ILS method can improve the classifier's performance with rank-1 of 94% along with mAP of 90.12% on PKU-Sketch-ReID and rank-1 of 33.42% along with mAP of 29.72% on MaSk1K dataset.Further GradCAM analysis shows that the classifiers trained using ILS use more appropriate features compared with non-ILS classifiers.