Cluster Contrast for Unsupervised Person Re-Identification

Zuozhuo Dai, Guangyuan Wang, Weihao Yuan, Siyu Zhu, Ping Tan; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 1142-1160

Thanks to the recent research development in contrastive learning, the gap of visual representation learning between supervised and unsupervised approaches has been gradually closed in the tasks of computer vision. In this paper, we focus on the downstream task of unsupervised person re-identification (re-ID). State-of-the-art unsupervised re-ID methods train the neural networks using a dictionary-based non-parametric softmax loss. They store the pre-computed instance feature vectors inside the dictionary, assign pseudo labels to them using clustering algorithm, and compare the query instances to the cluster using a form of contrastive loss. To enforce a consistent dictionary, that is the features in the dictionary are computed by a similar or the same encoder network, we present Cluster Contrast which stores feature vectors and computes contrastive loss at the cluster level. Moreover, the momentum update is introduced to reinforce the cluster-level feature consistency in the sequential space. Despite the straightforward design, experiments on four representative re-ID benchmarks demonstrate the effective performance of our method.