High-Confidence Sample Labelling for Unsupervised Person Re-identification

Person re-identification (re-ID) is factually a topic of pedestrian retrieval across camera scenes. However, it is challenging due to those factors such as complex equipment modeling, light change and occlusion. Much of the previous research is based on supervised methods that require labeling large amounts of data, which is expensive and time-consuming. The unsupervised re-ID methods without manual annotation usually need to construct pseudo-labels through clustering. However, the pseudo-labels noise may seriously affect the model’s performance. To deal with this issue, in this paper, we use Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to assign pseudo-labels to samples and propose a model with the high-confidence samples’ labels (HCSL), which is a fully unsupervised learning method and does not use any labeled data. The model constructs high-confidence triplets through cyclic consistency and random image transformation, which reduces noise and makes the model finely distinguish the differences between classes. Experimental results show that the performance of our method on both Market-1501 and DukeMTMC-reID performs better than the latest unsupervised re-ID methods and even surpasses some unsupervised domain adaptation methods.