Cluster-Guided Asymmetric Contrastive Learning for Unsupervised Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised person re-identification (Re-ID) seeks to match pedestrians across camera views without labeled data.
• Current Re-ID methods often rely on clustering, but feature quality is limited by color dominance.
• Developing robust features independent of color is crucial for effective unsupervised Re-ID.

Purpose of the Study:

• To propose a novel approach that mitigates the negative impact of color dominance in unsupervised person Re-ID.
• To enhance the learning of discriminant features for improved pedestrian matching accuracy.
• To introduce a Cluster-guided Asymmetric Contrastive Learning (CACL) framework.

Main Methods:

• Proposed the Cluster-guided Asymmetric Contrastive Learning (CACL) framework for unsupervised person Re-ID.
• Leveraged clustering results to guide feature learning within an asymmetric contrastive learning setup.
• Employed both instance-level and cluster-level contrastive learning with a siamese network.
• Introduced and validated a cluster refinement method to further improve performance.

Main Results:

• Demonstrated superior performance of the CACL approach on three benchmark datasets.
• Showcased the effectiveness of suppressing color dominance for learning better features.
• Validated the significant contribution of the cluster refinement step to CACL's performance.

Conclusions:

• The proposed CACL method significantly advances unsupervised person Re-ID by learning more effective features.
• Reducing color dependency in feature extraction leads to improved pedestrian matching.
• The integration of clustering guidance and asymmetric contrastive learning offers a promising direction for future Re-ID research.