Two-Stage Clustering Pseudo-Labels Correction for Cross-Domain Person Re-Identification

Abstract

The same identity of person images captured by multiple cameras may encounter appearance change significantly. In this case, most cluster-based cross-domain person re-identification(re-id) methods easily group them into different groups, which may generate lots of noise pseudo labels and limit the model performance. In this paper, we propose a Two-stage Clustering Pseudo-Labels Correction method (TCPC) based on camera ID to obtain reliable pseudo labels list. Specifically, we first cluster the samples of the same camera individually and assign them local pseudo-labels. Then, we cluster all samples to obtain global pseudo-labels. Using reliable local pseudo-labels to correct the global pseudo-labels, the impact of noisy pseudo-labels on model performance is reduced. Furthermore, we enhance the discriminativeness of pedestrian features by introducing the attention module to promote the clustering effect. We constructed a baseline and conducted comprehensive experiments on three widely used datasets, DukeMTMC-ReID, Market1501 and MSMT17, showing that our method improves the performance of the baseline. After that, we embedded the method into other the state-of-the-art frameworks to further verify its effectiveness.