Hybrid Contrastive Learning with Attention Mechanism for Unsupervised Person Re-identification

At present, learning discriminative features is crucial for unsupervised person re-identification model. Therefore, we introduce attention mechanism in the feature extraction stage to achieve more accurate clustering results. In addition, using only one feature to represent the whole cluster and update the memory may introduce bias which have adverse effect on the final result. So this paper adopt the memory bank updates based on average features and memory bank updates based on hard samples. Besides, we construct a contrastive learning loss function to ensure the stability in network updating. Experiments on two widely used public datasets of re-ID show that our approach can effectively improve the performance of unsupervised person re-identification.