Confidence-guided Outlier Refinement and Collaborative Embedding for Unsupervised Person Re-Identification

Unsupervised person re-identification techniques have developed rapidly in recent years. Nonetheless, they still face challenges such as unstable pseudo-label quality and insufficient feature representation, particularly when handling outlier points and complex backgrounds. To address these issues, this paper proposes a joint optimization algorithm of Multi-level Confidence Outlier Refinement (MLCOR) and Collaborative Embedding Method (CEM), which aims to improve the discriminative nature of the embedding space and optimize the accuracy of pseudo-labels. Specifically, Multi-level Confidence Outlier Refinement evaluates the confidence level of outlier points by analyzing the distance relationship between outlier points and their neighboring samples, and classifies them into multiple confidence levels. We design a weighted voting strategy for low-confidence samples to correct the pseudo-labeling of low-confidence points by using the label distribution of neighboring samples, thus reducing noise interference and clustering errors and improving the accuracy of pseudo-labeling. Meanwhile, the Collaborative Embedding Method jointly optimizes global and local features, establishing an effective synergy between global category differentiation and local fine-grained feature learning. By integrating multi-level similarity relationships, this approach not only strengthens the model’s ability to capture subtle differences between samples but also significantly enhances the model's boundary awareness. Experimental results demonstrate that the proposed method achieves outstanding performance on multiple standard datasets, significantly improving both clustering accuracy and pseudo-label precision, while also exhibiting strong domain generalization and robustness in complex environments.