Unsupervised person reidentification via quantitative random selection for cluster centroid

Person reidentification (re-ID) is an important topic in computer vision. This paper studies an unsupervised approach to re-ID, which does not require any labeled information and is thus possible to deploy in real-world scenarios. State-of-the-art unsupervised re-ID methods usually use a memory bank to store the instance feature vectors, generate pseudolabels with a clustering algorithm, and compare the query instances to the centroid of the clusters for contrastive learning. However, because hard negative or noisy samples exist, the centroid generated by unsupervised learning may not be a perfect prototype. Forcing the wrong images to get closer to the centroid would result in accumulated errors and deteriorated overfitting. To solve this problem, we propose a quantitative random selection strategy to form the cluster feature representation. Specifically, in each iteration, the cluster algorithm executes on instance-level feature vectors to generate pseudolabels. Then, we shuffle all the instance vectors belonging to the same cluster and select samples within the same cluster in a certain proportion to form the cluster-level memory. During network training, the query instances are used to update the cluster-level memory for contrastive learning. Extensive experiments show that our proposed method produces state-of-the-art performance in unsupervised person re-ID tasks.

Automated summary

This paper studies an unsupervised approach to person reidentification, using a clustering algorithm and contrastive learning to generate pseudolabels, and proposes a quantitative random selection strategy for cluster feature representation. Extensive experiments show that this method achieves state-of-the-art performance in unsupervised person re-ID tasks.