Cross-Modality Person Re-Identification via Local Paired Graph Attention Network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Cross-modality person re-identification (ReID) addresses the challenge of matching pedestrian images between different visual spectra, such as RGB and infrared (IR).
• Existing graph-based methods for cross-modality ReID often overlook the inherent correlations within paired IR and RGB images, potentially limiting performance.
• Bridging the modality gap between RGB and IR imagery remains a significant challenge in computer vision.

Purpose of the Study:

• To propose a novel graph model, the Local Paired Graph Attention Network (LPGAT), for enhancing cross-modality person re-identification.
• To address the limitations of existing methods by incorporating paired local features and contextual information.
• To improve the accuracy of person re-identification across different modalities by learning a more robust distance metric.

Main Methods:

• Development of the Local Paired Graph Attention Network (LPGAT) model, utilizing paired local features from different modalities as graph nodes.
• Introduction of a contextual attention coefficient that incorporates distance information to refine information propagation within the graph.
• Implementation of Cross-Center Contrastive Learning (C3L) to optimize the learning of the distance metric by constraining feature distributions relative to heterogeneous centers.

Main Results:

• Experimental validation on the RegDB and SYSU-MM01 datasets demonstrates the effectiveness of the proposed LPGAT approach.
• The method successfully leverages paired local features and attention mechanisms to improve cross-modality matching accuracy.
• Cross-Center Contrastive Learning contributes to learning a more discriminative feature representation for ReID.

Conclusions:

• The proposed LPGAT model, incorporating paired local features and contextual attention, offers a significant advancement in cross-modality person re-identification.
• The integration of Cross-Center Contrastive Learning further enhances the model's ability to learn discriminative features across modalities.
• The approach provides a feasible and effective solution for matching pedestrians between RGB and IR imagery.