Confidence-guided outlier refinement and collaborative embedding for unsupervised person re-identification.

Area of Science:

• Computer Science
• Artificial Intelligence

Background:

• Unsupervised person re-identification faces challenges with pseudo-label quality and feature representation.
• Outlier points and complex backgrounds degrade performance in existing methods.

Purpose of the Study:

• To enhance unsupervised person re-identification by improving pseudo-label accuracy and feature representation.
• To develop a robust method for handling complex backgrounds and outlier data.

Main Methods:

• Proposes a joint optimization algorithm combining Multi-level Confidence Outlier Refinement (MLCOR) and Collaborative Embedding Method (CEM).
• MLCOR refines pseudo-labels by analyzing outlier confidence and using weighted voting for low-confidence samples.
• CEM jointly optimizes global and local features, integrating multi-level similarity for enhanced discrimination and boundary awareness.

Main Results:

• Achieves outstanding performance on standard person re-identification datasets.
• Significantly improves clustering accuracy and pseudo-label precision.
• Demonstrates strong domain generalization and robustness in complex environments.

Conclusions:

• The proposed MLCOR and CEM algorithm effectively addresses limitations in unsupervised person re-identification.
• The method enhances discriminative embedding and pseudo-label accuracy, leading to superior performance.