FA-Net: A Feature Alignment Network for Video-Based Visible-Infrared Person Re-Identification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video-based visible-infrared person re-identification (VVI-ReID) is crucial for 24-hour surveillance.
• Existing methods struggle with temporal misalignment and domain shift noise, focusing mainly on modality differences.

Purpose of the Study:

• To propose a novel VVI-ReID framework, FA-Net, that mitigates temporal misalignment and domain shift noise.
• To enhance sequence-level representation learning for improved cross-modality pedestrian matching.

Main Methods:

• Introduced FA-Net with Spatial-Temporal Alignment Module (STAM) for spatial and temporal feature alignment.
• Employed Modality Distribution Constraint (MDC) using symmetric distribution loss for feature distribution alignment.
• Utilized SAM Guidance Augmentation (SAM-GA) to enhance frame information quality.

Main Results:

• FA-Net effectively addresses temporal misalignment and domain shift noise in VVI-ReID.
• The proposed method surpasses existing state-of-the-art VVI-ReID techniques.
• Experimental results validate the framework's superior performance.

Conclusions:

• FA-Net offers a significant advancement in VVI-ReID by focusing on feature alignment.
• The framework improves the robustness and accuracy of pedestrian re-identification in surveillance systems.
• The developed approach contributes to more effective 24-hour surveillance solutions.