A new robust contrastive learning for unsupervised person re-identification

Unsupervised person re-identification (Re-ID) is more substantial than the supervised one because it does not require any labeled samples. Currently, the most advanced unsupervised Re-ID models generate pseudo-labels to group images into different clusters and then establish a memory bank to calculate contrastive loss between instances and clusters. This framework has been proven to be remarkably efficient for unsupervised person Re-ID tasks. However, clustering operation inevitably produces misclassification, which brings noises and difficulties to contrastive learning and affects the initialization and updating of the prototype features stored in the memory bank. To solve this problem, we propose a new robust unsupervised person Re-ID model with two developed modules: Cluster Sample Aggregation module (CSA) and Hard Positive Sampling strategy (HPS). The CSA module aggregates each sample in the same cluster through the multi-head self-attention mechanism. This process enables the initialization of prototypes based on the similarities observed within clusters. Additionally, the HPS strategy extracts the dispersion degree of each sample by means of a self-attention aggregation module (SAA) that has been trained by CSA module. According to the obtained indicators, the hardest positive sample is sampled to update the prototype feature stored in the memory bank. With the self-attention mechanism fusing the information among instances in each cluster, the implicit relationships between samples can be better explored in a more refined way. Experiments show that our method achieves state-of-the-art results against existing unsupervised baselines on Market-1501, PersonX, and MSMT17 datasets.

Automated summary

This study proposes a new robust unsupervised person re-identification model that solves the problem of misclassification caused by clustering operations, achieving a better exploration of implicit relationships between samples.