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

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

Problem Statement
Remote Photoplethysmography (rPPG) enables contactless estimation of vital signs, such as heart rate, by analyzing subtle color and intensity variations in facial videos. However, current deep learning methods struggle in complex, unconstrained scenarios characterized by varying illumination, facial expressions, and head movements. In these environments, subtle physiological signals are often overwhelmed by external interference. The authors identify a primary limitation in existing approaches: most methods model the rPPG signal in a unified manner, where different signal components (e.g., pulse-related variations and disturbance-induced changes) are coupled during reconstruction. This coupling makes it difficult to preserve weak pulse-related variations when strong disturbances are present. Furthermore, existing methods often rely on iterative denoising (e.g., diffusion models), leading to high computational latency, and frequently prioritize heart rate (HR) accuracy over the morphological fidelity of the reconstructed waveform.

Methodology: PhysFlow
To address these challenges, the authors propose PhysFlow, a frequency-decoupled dual-field rectified flow framework. The core innovation lies in separating the modeling of the rPPG signal into distinct components rather than treating it as a unified whole.

1. Frequency-Domain Decomposition (FDD):
   The ground-truth rPPG signal is decomposed into two distinct supervisory targets:

   • Trend Component: Captures the slowly varying waveform profile (low-frequency).
   • Amplitude Component: Preserves fine-grained, pulse-related variations within the plausible heart-rate range.
     This decomposition is achieved using a real Fast Fourier Transform (RFFT) with specific low-frequency and band-pass masks, followed by an inverse transform (IRFFT).

2. Dual-Field Rectified Flow Framework:
   Instead of using these components merely as auxiliary cues, PhysFlow treats them as explicit generation targets. The framework learns two separate conditional velocity fields:

   • Trend Head: Designed to model the smooth, low-frequency trend. It employs a basis reconstruction approach (using spectral and polynomial bases) to enforce smoothness and suppress unnecessary high-frequency fluctuations.
   • Amplitude Head: Designed to model the pulse-related variations, which are more sensitive to video-dependent disturbances. It utilizes a condition-aware temporal predictor incorporating FiLM modulation, dilated temporal convolutions, and Squeeze-and-Excitation (SE) reweighting to capture complex temporal dependencies.
     Both heads are guided by features extracted from the input video via a Periodic Transformer Block, which includes a global feature shared by both branches and component-specific guidance features.

3. Efficient Inference via Rectified Flow:
   Unlike traditional diffusion models that require iterative denoising, PhysFlow utilizes the Rectified Flow formulation. It learns a velocity field to transport samples from a Gaussian-initialized latent state directly to the target trend and amplitude states along a straight-line path. During inference, the final rPPG signal is reconstructed by integrating ordinary differential equations (ODEs) with only a few steps (e.g., 4 steps), significantly reducing computational overhead.

Key Contributions

• Dual-Field Architecture: The paper proposes a novel framework that decouples the trend and amplitude components of the rPPG signal, modeling them via two distinct conditional velocity fields to reduce mutual interference.
• Explicit Supervisory Decomposition: A frequency-domain decomposition strategy is introduced to derive explicit supervisory targets for the trend and amplitude components, allowing the decomposition to actively participate in the generative process rather than serving as a post-hoc aid.
• Efficient and Robust Reconstruction: By leveraging rectified flow, the method achieves high-quality waveform reconstruction with minimal ODE integration steps, addressing the latency issues of diffusion-based approaches while improving robustness against complex disturbances.

Experimental Results
The authors evaluated PhysFlow on four benchmark datasets: BUAA-MIHR (low illumination), VIPL-HR (diverse motions and lighting), NIRP-DRV (driving scenarios), and NIRP-IND (indoor driving).

• Heart Rate Estimation: PhysFlow achieved state-of-the-art performance across all datasets. For instance, on the challenging NIRP-DRV dataset, it achieved an RMSE of 5.93 bpm, outperforming the previous best (PHASE-Net) of 6.23 bpm. On the BUAA-MIHR dataset, it reduced RMSE to 1.17 bpm compared to 1.43 bpm for the runner-up.
• Waveform Reconstruction: Beyond HR accuracy, PhysFlow demonstrated superior waveform fidelity. On the BUAA dataset, it achieved a Waveform Correlation (Wave-Corr) of 0.95, a Phase Error (PE) of 0.12, and an SNR of 17.18, significantly outperforming other methods in preserving morphology and spectral quality.
• Cross-Dataset Generalization: In cross-dataset evaluations (e.g., training on NIRP and testing on VIPL-HR), PhysFlow maintained the best overall performance, indicating superior generalization capabilities across different acquisition devices and environments.
• Efficiency: Despite not having the smallest parameter count, PhysFlow demonstrated the highest throughput (63.74 Kps) and lowest inference time (13.63 ms) on an NVIDIA RTX 4090 GPU among compared methods, validating its suitability for real-time deployment.

Significance and Claims
The paper claims that PhysFlow represents a significant advancement in rPPG estimation by explicitly addressing the entanglement of physiological signals and environmental disturbances. By separating the signal into trend and amplitude components and modeling them with specialized velocity fields, the method improves robustness without sacrificing efficiency. The authors emphasize that accurate HR estimation does not guarantee faithful waveform recovery; therefore, their focus on morphological fidelity provides a more comprehensive evaluation of rPPG quality.

The authors modestly acknowledge limitations, noting that the current model complexity may hinder lightweight deployment and that the fixed frequency cutoffs in the decomposition strategy might limit flexibility under abnormal heart rates or extreme motion. Future work is proposed to develop lighter feature extraction frameworks and adaptive decomposition strategies.