Contrast-Phys+: Unsupervised and Weakly-Supervised Video-Based Remote Physiological Measurement via Spatiotemporal...

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote photoplethysmography (rPPG) measures physiological signals from facial videos.
• Supervised methods achieve good performance but require costly ground truth (GT) data.
• Acquiring accurate GT physiological signals for rPPG training is challenging.

Purpose of the Study:

• To develop a robust rPPG measurement method trainable in unsupervised and weakly-supervised settings.
• To overcome the limitations of supervised methods requiring extensive GT data.
• To improve the efficiency, robustness, and generalization of rPPG estimation.

Main Methods:

• Proposed Contrast-Phys+ method utilizing a 3DCNN model for spatiotemporal rPPG signal generation.
• Incorporated prior rPPG knowledge into a contrastive loss function.
• Integrated GT signals into contrastive learning to handle partial or misaligned labels.

Main Results:

• Contrast-Phys+ outperformed state-of-the-art supervised methods on five diverse datasets (RGB and Near-infrared).
• The method demonstrated strong performance with partially available, misaligned, or absent GT signals.
• Evaluated advantages in computational efficiency, noise robustness, and generalization capabilities.

Conclusions:

• Contrast-Phys+ offers a flexible and effective solution for rPPG measurement, reducing reliance on perfect GT data.
• The proposed contrastive learning approach enhances rPPG estimation accuracy and applicability.
• This method presents a significant advancement for non-contact physiological monitoring.