Graph Sampling-Based Multi-Stream Enhancement Network for Visible-Infrared Person Re-Identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Single-modal person re-identification (Re-ID) is insufficient for all-day retrieval needs.
• Multi-modal data, specifically Visible-Infrared Person Re-Identification (VI Re-ID), is crucial but challenged by significant modality discrepancies.
• Existing methods often overlook the modality-invariant nature of person contour information.

Purpose of the Study:

• To address the challenges in VI Re-ID, particularly modality discrepancy and distinguishing hard samples.
• To propose a novel network that leverages contour information and advanced sampling techniques for improved cross-modal matching.
• To enhance the stability and accuracy of person matching across visible and infrared modalities.

Main Methods:

• Proposed the Graph Sampling-based Multi-stream Enhancement Network (GSMEN) incorporating a Contour Expansion Module (CEM).
• CEM integrates person contour information to reduce modality discrepancy and enhance matching stability.
• Introduced a Cross-modality Graph Sampler (CGS) for intelligent sample selection, grouping similar cross-modal samples to explore hard class boundaries.

Main Results:

• Experiments on SYSU-MM01 and RegDB datasets demonstrate the superiority of the proposed GSMEN.
• Achieved 93.69% Rank-1 and 92.56% mAP in the VIS→IR task on the RegDB dataset.
• The method effectively reduces modality discrepancy and improves the distinction of hard cross-modal samples.

Conclusions:

• The proposed GSMEN effectively tackles the modality discrepancy challenge in VI Re-ID by utilizing contour information.
• The Cross-modality Graph Sampler aids in training by focusing on difficult cross-modal sample pairs.
• The method shows significant performance improvements, highlighting the importance of contour and intelligent sampling in VI Re-ID.