Unsupervised Long-Term Person Re-Identification with Clothes Change

Most person re-identification methods artificially assume that each person's clothing is stationary in space and time. Since the average person often changes clothes even within a single day, this condition primarily holds true in situations involving short-term re-identification scenarios. Some recent studies have investigated re-identification of clothing changes based on supervised learning to reduce this limitation. In this paper, we remove the necessity for personal identity labels, which makes this new problem dramatically more challenging than conventional unsupervised short-term Re-ID. To surmount these obstacles, we introduce a novel approach known as the Curriculum Person Clustering (CPC) method, which exhibits the ability to dynamically modify the clustering criterion based on the clustering confidence in the clustering process. Experimental results on DeepChange show that CPC surpasses other unsupervised re-id method and even close to supervised methods.