Hierarchical Identity Learning for Unsupervised Visible-Infrared Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised visible-infrared person reidentification (USVI-ReID) aims to bridge the modality gap in unlabeled datasets.
• Current cluster-based contrastive learning methods often overlook fine-grained variations within person identity clusters.
• Reliance on manual annotations is a significant limitation in existing USVI-ReID approaches.

Purpose of the Study:

• To propose a Hierarchical Identity Learning (HIL) framework to address limitations in USVI-ReID.
• To improve the learning of modality-invariant features by considering finer-grained differences.
• To reduce the need for manual annotations in person reidentification tasks.

Main Methods:

• Implemented a Hierarchical Identity Learning (HIL) framework utilizing secondary clustering to generate multiple memories per coarse-grained cluster.
• Introduced Multi-Center Contrastive Learning (MCCL) to refine representations, enhance intra-modal clustering, and minimize cross-modal discrepancies.
• Developed a Bidirectional Reverse Selection Transmission (BRST) mechanism for reliable cross-modal correspondences through bidirectional pseudo-label matching.

Main Results:

• The proposed HIL framework demonstrated superior performance on the SYSU-MM01 and RegDB datasets.
• MCCL effectively refined representations and improved both intra-modal clustering and cross-modal matching.
• BRST mechanism successfully established reliable cross-modal correspondences, enhancing overall reidentification accuracy.

Conclusions:

• The Hierarchical Identity Learning (HIL) framework offers a significant advancement in unsupervised visible-infrared person reidentification.
• The proposed methods effectively capture fine-grained identity variations and reduce the modality gap.
• Experimental results validate the effectiveness and superiority of the HIL framework over existing approaches.