PhysFlow: Frequency Decoupled with Dual-Field Rectified Flow for Remote Photoplethysmography

2026-06-22 · Source: Takara TLDR - Daily AI Papers · Field: Science & Research — Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, medium

Summary

PhysFlow is a novel frequency-decoupled dual-field rectified flow framework designed to enhance robust remote Photoplethysmography (rPPG) estimation from facial videos. Current deep learning rPPG methods struggle with complex disturbances like varying illumination, facial expressions, and head movements, which often mask subtle physiological signals. PhysFlow addresses this by decomposing the ground-truth rPPG signal into separate trend and amplitude components, using them as distinct supervisory targets. It then learns two component-specific conditional velocity fields from extracted facial features, reducing mutual interference and improving rPPG reconstruction stability. The rectified flow formulation further enables efficient waveform reconstruction with only a few ordinary differential equation (ODE) integration steps. Extensive experiments on multiple benchmark datasets demonstrate PhysFlow's superior performance in both heart-rate estimation and rPPG waveform reconstruction across diverse challenging scenarios compared to existing advanced methods.

Key takeaway

For Machine Learning Engineers developing contactless health monitoring systems, PhysFlow offers a robust solution for rPPG estimation. You should consider implementing its frequency-decoupled, dual-field rectified flow approach to improve accuracy under varying illumination and head movements. This method enhances signal stability and reliability, crucial for real-world applications, by separately modeling physiological signal components.

Key insights

PhysFlow improves rPPG robustness by frequency-decoupling signal components and modeling them separately with rectified flow.

Principles

• Decouple complex signals into simpler components.
• Model distinct signal components independently.
• Rectified flow enables efficient ODE integration.

Method

Decompose rPPG into trend and amplitude components. Learn component-specific conditional velocity fields. Reconstruct waveforms using few ODE integration steps.

In practice

• Enhance contactless health monitoring.
• Improve rPPG in challenging environments.
• Apply to other physiological signal analysis.

Topics

• Remote Photoplethysmography
• Rectified Flow
• Physiological Signal Processing
• Deep Learning Models
• Health Monitoring
• Signal Decomposition

Code references

• G-U-N/Rectified-Diffusion
• Xinxi-Zhang/Re-MeanFlow
• Harvard-AI-and-Robotics-Lab/DeltaRectifiedFlowSampling
• aaFrostnova/Papillon

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• StreamPPG: Low-Latency rPPG Estimation via Consistent Privileged Learning — StreamPPG leverages consistent privileged learning to enable low-latency, accurate remote photoplethysmography estimation suitable for real-time edge device deployment.
• RelativeFlow: Taming Medical Image Denoising Learning with Noisy Reference — RelativeFlow uses flow matching to learn unified denoising from heterogeneous noisy medical image references.
• Intervention-Based Self-Supervised Learning: A Causal Probe Paradigm for Remote Photoplethysmography — Intervention-based self-supervised learning can overcome correlation traps in rPPG by actively verifying physical signal hypotheses.
• Google: Detecting Heart Rate Through Smartphone Camera AI — Smartphone cameras can passively monitor heart rate with medical-grade accuracy across diverse skin tones using deep learning.
• Thermal Imaging for Contactless Cardiorespiratory and Sudomotor Response Monitoring — Thermal imaging can non-invasively monitor cardiorespiratory and sudomotor responses, with varying accuracy based on signal type and camera frame rate.
• DreamReg: Belief-Driven World Model for 2D-3D Ultrasound Registration — DreamReg uses a belief-driven world model and internal imagination for robust 2D-3D ultrasound registration, overcoming real-time challenges.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.