Enhancing Unsupervised Domain Adaptive Person Re-identification Clustering Through Parsing-Based Attribute Labeling

Cross-domain person re-identification (re-ID) is an essential component of industrial video surveillance systems because it is capable of retrieving the target identity across multiple non-overlapping camera circumstances. While clustering-based methods are widely adopted in person re-ID and deliver significant success, these algorithms exhibit certain limitations, i.e., 1) The high dependency on the quality of clustering results, which severely affects the model’s performance; 2) Poor initial model performance can result in suboptimal clustering outcomes and foster a detrimental feedback loop during training. To solve the aforementioned issues, this paper introduces image-level attribute labels for pedestrian samples, which serve a dual purpose: providing semantic information to aid the model in focusing on the human body and reducing the interference from background noise. To make it clean, a novel Auxiliary Attribute Clustering (AAC) pipeline for unsupervised domain adaptive person re-ID is proposed to integrate the visual and semantic information. Extensive experiments on three challenging person re-ID datasets are conducted to demonstrate the promising performance of the proposed method.