Mask-Guided Asymmetric Contrastive and Semantic Alignment for Unsupervised Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised person re-identification (ReID) faces challenges like noisy labels and appearance variations.
• Exploiting fine-grained local cues and random masking are key strategies in unsupervised ReID.
• Existing methods underutilize masked views and have weak cross-view alignment.

Purpose of the Study:

• To propose a novel framework, Mask-guided Asymmetric Contrastive and Semantic Alignment (ACSA), to address limitations in unsupervised ReID.
• To enhance the utilization of masked views as supervisory signals and improve global feature alignment.
• To achieve more robust and accurate identity representations in unsupervised ReID.

Main Methods:

• Introduced an Asymmetric Contrastive Learning (ACL) module with dual-memory for separate encoding of masked and unmasked features.
• Developed a Semantic Alignment Learning (SAL) module for multi-granularity distribution alignment (cluster and instance level).
• Incorporated a Progressive Refinement Module (PRM) to refine prototypes and features for stable semantic alignment.

Main Results:

• The proposed ACSA framework effectively utilizes masked views as discriminative supervision.
• ACSA achieves superior performance in unsupervised person ReID, outperforming existing methods.
• Experimental validation demonstrates the method's superiority, even surpassing some supervised approaches.

Conclusions:

• ACSA framework significantly advances unsupervised person ReID by addressing key limitations in feature learning and alignment.
• The combination of asymmetric contrastive learning and multi-granularity semantic alignment proves highly effective.
• The method offers a robust solution for learning identity-discriminative representations without manual annotations.