Clothing-invariant contrastive learning for unsupervised person re-identification

Zhiqi Pang, Lingling Zhao, Chunyu Wang

Abstract: Highlights•Clothing change positive pairs and identity-related pseudo-labels are essential.•Random clothing augmentation efficiently generates clothing change positive pairs.•Semantic fusion clustering enhances identity-related information in the features.•The novel loss improves the model’s robustness to identity-irrelevant attributes.