Attend to the Difference: Cross-Modality Person Re-Identification via Contrastive Correlation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Cross-modality person re-identification (ReID) is crucial for surveillance and security.
• Existing methods struggle with capturing subtle differences between modalities like RGB and Infrared (IR).

Purpose of the Study:

• To develop a novel dual-path framework for effective cross-modality feature learning.
• To preserve spatial structures while emphasizing discriminative features in image pairs.

Main Methods:

• A Dual-path Spatial-structure-preserving Common Space Network (DSCSN) embeds images into a shared 3D tensor space.
• A Contrastive Correlation Network (CCN) extracts features by dynamically comparing image pairs, ensuring mutual dependency.
• The framework learns from both RGB and Infrared image data.

Main Results:

• The proposed method achieves state-of-the-art performance on the SYSU-MM01 and RegDB datasets.
• Significant improvements were observed in both full and simplified evaluation modes.
• The dual-path approach effectively captures cross-modality differences and preserves spatial information.

Conclusions:

• The dual-path framework offers a superior approach to cross-modality person ReID.
• Highlighting differences and preserving spatial structures are key to robust ReID performance.
• This work advances the capabilities of RGB-IR person re-identification systems.