Learning to Purification for Unsupervised Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Unsupervised person re-identification (ReID) methods often struggle with feature and label noise from pseudo-label training.
• Existing approaches lack explicit strategies for purifying these noisy elements in an unsupervised manner.

Purpose of the Study:

• To develop effective purification modules for unsupervised person ReID.
• To address the challenges of feature representation enrichment and label noise reduction.

Main Methods:

• Multi-view feature integration to enrich representation and capture discriminative cues.
• A teacher-student model scheme for label noise purification using an offline approach.
• Cluster contrast learning framework to leverage purified features and labels.

Main Results:

• The proposed purification modules significantly improve unsupervised person ReID performance.
• Achieved state-of-the-art accuracy of 85.8% @mAP and 94.5% @Rank-1 on the Market-1501 dataset using ResNet-50.
• Demonstrated effectiveness across multiple benchmark datasets.

Conclusions:

• The developed purification strategy effectively handles noise and bias in unsupervised person ReID.
• The multi-view feature enrichment and teacher-guided label purification are key to the method's success.
• The approach offers a robust solution for challenging unsupervised ReID tasks.