Generalizable Remote Physiological Measurement via Semantic-Sheltered Alignment and Plausible Style Randomization

Jiyao Wang, Hao Lu, Hu Han, Yingcong Chen, Dengbo He, Kaishun Wu

Abstract: Remote photoplethysmography (rPPG) is a noninvasive technique that measures blood volume changes in the skin using a camera and a light source. Achieving accurate measurements relies on the generalization of models across different individuals and environmental conditions. However, most domain generalization (DG) methods are designed for classification tasks rather than regression tasks, which is a suboptimal solution for the rPPG task. In this work, we propose a novel dual-stream generalization framework (DG-rPPG), which consists of semantic-sheltered alignment (SSA) and plausible attribute randomization (PAR). Specifically, SSA can extract and align domain-agnostic features from different datasets; while maximumly preserving semantic information. PAR can enrich the attribute-related feature of each instance based on the statistical information of all the different domains, ensuring that the augmented features maintain plausibility. The heart rate (HR) and HR variability estimation evaluation with cross-domain protocol across five public datasets illustrated that our proposal significantly (p-value <0.05) outperforms all baselines (e.g., compared to DOHA, DG-rPPG achieves 9.25% and 8.19% improvement on MAE when UBFC and BUAA are target domains). Meanwhile, based on the intra-dataset, computation cost, and out-of-distribution (OOD) assessment, DG-rPPG presents the leading performance in OOD generalization while maintaining relatively good performance in in-distribution estimation and reasonable computational costs. This provides a foundation for real-time monitoring deployments in real environments. The code is available at https://github.com/WJULYW/DG-rPPG.