Hierarchical Style-Aware Domain Generalization for Remote Physiological Measurement.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote photoplethysmography (rPPG) enables non-contact measurement of blood volume pulse (BVP) from facial videos.
• Existing rPPG methods face challenges in generalizing to unseen domains due to environmental and individual variations.
• Domain generalization in rPPG is hindered by ambiguous instance-specific differences, limiting current approaches.

Purpose of the Study:

• To address the domain shift problem in rPPG measurement.
• To develop a novel method that improves the generalization capacity of rPPG.
• To enhance the robustness of rPPG analysis for out-of-distribution samples.

Main Methods:

• Proposes Hierarchical Style-aware Representation Disentangling (HSRD) for rPPG.
• HSRD separates domain-invariant and instance-specific feature spaces during training.
• Improves generalization by disentangling features to increase robustness.

Main Results:

• HSRD demonstrates superior performance compared to existing methods.
• Achieves state-of-the-art results in both cross-dataset and intra-dataset settings.
• Shows improved robustness for out-of-distribution samples during inference.

Conclusions:

• HSRD effectively tackles the domain shift problem in rPPG.
• The proposed method enhances the generalization capability of rPPG technology.
• HSRD offers a more robust solution for health monitoring and emotional analysis applications.