Channel semantic mutual learning for visible-thermal person re-identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Visible-infrared person re-identification (VI-ReID) is crucial for surveillance and security.
• Modality discrepancy between visible and infrared imagery poses a significant challenge.
• Existing methods often use separate networks, limiting cross-modality feature learning.

Purpose of the Study:

• To propose a novel Channel Semantic Mutual Learning Network (CSMN) for VI-ReID.
• To address the modality discrepancy by focusing on channel-level semantic differences.
• To improve the accuracy and robustness of pedestrian matching across different imaging modalities.

Main Methods:

• Developed CSMN, a network that optimizes semantic consistency between channels from local and global perspectives.
• Introduced a Channel-level Auto-Guided Double Metric loss (CADM) for fine-grained learning of modality-invariant features.
• Employed dual optimization strategies for enhanced semantic alignment.

Main Results:

• CSMN demonstrated superior performance on the RegDB and SYSU-MM01 datasets.
• Achieved a 3.43% improvement in Rank-1 score and a 0.5% improvement in mINP on the RegDB dataset.
• Validated the effectiveness of channel-level semantic mutual learning and the CADM loss.

Conclusions:

• The proposed CSMN effectively tackles the modality discrepancy in VI-ReID.
• Channel-level semantic mutual learning offers a promising direction for cross-modality retrieval.
• CSMN sets a new state-of-the-art performance benchmark for VI-ReID tasks.