Transfer easy to hard: Adversarial contrastive feature learning for unsupervised person re-identification

Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, Gang Hua

Abstract: Highlights•We design a novel adversarial contrastive feature learning (ACFL) framework for unsupervised person Re-ID, which can generate hard samples with high-confidence pseudo labels to guide the discriminative feature learning process.•We design a discriminative feature learning (DFL) module to incorporate these hard samples to further enhance the discriminative capability of features.•We design a novel hard sample generation (HSG) module to generate hard samples for discriminative feature learning, in which an adversarial learning regime is used to generate hard person features (class-level) and hard positive instances (instance-level).