Uncertainty Quantification for Deep Learning-Based Remote Photoplethysmography
Rencheng Song, Han Wang, Haojie Xia, Juan Cheng, Chang Li, Xun Chen
Abstract
Recently, deep learning (DL) based remote photoplethysmography (rPPG) methods have made great progress in non-contact heart rate (HR) measurements. However, the reliability of rPPG measurements is still a bottleneck for realistic applications due to its susceptibility to motion and illumination noise. It is increasingly important to develop quantitative assessment methods for rPPG measurements. In this paper, we propose to quantify the uncertainty of a convolutional neural network (CNN)-based rPPG model as the reference-free quantitative quality assessment metric. Specifically, we evaluate and compare both the Bayesian neural network (BNN) method and the deep ensemble method to estimate the uncertainty of HR measurements from the CNN model. The effectiveness of the predicted uncertainty from both methods is verified on two public databases, where the uncertainty is found to be highly correlated with the true absolute error, and the advantages and limitations of BNN and ensemble methods are also analyzed. This study is instructive for the development of uncertainty-aware DL-based rPPG models for realistic applications.