The Double Contrast for Unsupervised Person Re-Identification

At present, the most popular unsupervised person re-identification(Re-ID) research mainly uses some clustering methods to gather samples with similar features and generate a pseudo-label for each cluster, and uses the parameterless contrast loss to feedback correlation information between features.Many previous methods used the memory dictionary to store all instance-dependent features, but This approach does not take into account the different number of positive samples in each cluster, and the progress of each cluster update of the memory dictionary is not the same.In order to improve the consistency of the update progress, a method is implemented to store the central feature vector in each cluster, but using a hard-sample feature vector in the cluster to update it, so that a single instance feature cannot represent the central feature well. In this article, we study out a method to combine hard instance contrast and cluster contrast, that is, the central feature vector in each cluster is stored in the memory dictionary, and the hard instance feature vector in each cluster is also stored there.In addition, in terms of momentum update, the central feature vector and hard instance feature vector of each cluster in mini-batch are used for update.This not only solves the problem of inconsistent update progress, but also better learns the similarity and difference of instances in the same cluster, and better aggregates instances within cluster and instances between clusters. Through three dataset experiments, it is concluded that this method has a good improvement.