Abstract
Accelerated by telemedicine, advances in Remote Photoplethysmography (rPPG) are beginning to offer a viable path toward non-contact physiological measurement. Unfortunately, the datasets for rPPG are limited as they require videos of the human face paired with ground-truth, synchronized heart rate data from a medical-grade health monitor. Also troubling is that the datasets are not inclusive of diverse populations, i.e., current real rPPG facial video datasets are imbalanced in terms of races or skin tones, leading to accuracy disparities on different demographic groups. This paper proposes a scalable biophysical learning based method to generate physio-realistic synthetic rPPG videos given any reference image and target rPPG signal and shows that it could further improve the state-of-the-art physiological measurement and reduce the bias among different groups. We also collect the largest rPPG dataset of its kind (UCLA-rPPG) with a diverse presence of subject skin tones, in the hope that this could serve as a benchmark dataset for different skin tones in this area and ensure that advances of the technique can benefit all people for healthcare equity.
Introduction
Remote Photoplethysmography (rPPG) is the science of measuring vital signs—like heart rate—by detecting subtle color changes in a person's face using a standard camera. While promising for telemedicine, rPPG faces a massive "diversity crisis."
Existing datasets are small because they require synchronized, medical-grade ground truth data, which is difficult and expensive to collect. More importantly, these datasets are heavily biased toward lighter skin tones. Because rPPG relies on light reflecting off the skin, melanin acts as an "optical attenuator," making the signal much weaker and harder to detect in darker-skinned individuals. This leads to significant accuracy disparities across different demographic groups. This work addresses these challenges by introducing a physics-based synthetic generation pipeline and a new, diverse real-world benchmark.
Methods
The proposed cross-modal synthetic generation model can turn any static face image and any target PPG signal into a realistic video. The pipeline consists of several sophisticated steps:
- 3D Face Inference: A Non-linear 3D Morphable Model (DECA) extracts the 3D shape and a UV albedo map from a reference image.
- Biophysical Decomposition: Using a physics-based framework (BioFaceNet) grounded in Kubelka-Munk theory, the skin albedo is decomposed into melanin and hemoglobin (blood) maps.
- Temporal Modulation: A target PPG waveform modulates the hemoglobin map over time, simulating blood pulsing through facial tissue.
- Physiological Rendering: The modulated maps are recombined with the 3D mesh, realistic lighting (Spherical Harmonics), and randomized head motions.
Datasets and Diversity
The paper introduces two major resources to address data scarcity and demographic bias:
- UCLA-rPPG: A new real-world benchmark containing 102 subjects with a diverse range of skin tones (Fitzpatrick scales 1-6), totaling 503 synchronized videos.
- Synthetic Dataset: 480 synthetic videos generated using the proposed method, specifically designed to be demographically balanced.
Key Results
The integration of physics-based synthetic data yields significant improvements in both accuracy and fairness:
- Accuracy Boost: Adding synthetic data to the training of state-of-the-art models (like PhysNet) reduced the Mean Absolute Error (MAE) from 1.06 BPM to 0.71 BPM on the UCLA-rPPG dataset.
- Closing the Gap: The most dramatic improvements were seen in the "Dark Skin" group (Fitzpatrick 5-6), where the error was reduced by 41%.
- Bias Reduction: The disparity (standard deviation of error) between different skin tone groups was reduced by 45%, making the technology significantly more equitable.
- Cross-Dataset Success: On the UBFC-rPPG dataset, a model trained only on synthetic data outperformed models trained on other real-world datasets, highlighting the high fidelity of the synthetic pulse.
Discussion and Conclusion
This work demonstrates that physics-based "fake" data is a powerful tool for bias mitigation. Unlike standard GANs that might just learn to mimic visual patterns, this model uses the actual physics of light transport in skin. This ensures the synthetic "pulse" is physiologically meaningful, which helps neural networks learn the underlying biology rather than just noise.
While "physio-realistic," the videos are not yet perfectly photo-realistic, and future work will focus on reducing the "sim-to-real" gap in facial textures. Nonetheless, the UCLA-rPPG dataset and the synthetic generation pipeline represent a major step toward healthcare equity in remote physiological monitoring.
Summary updated