Bayesian Optimization for Realistic PPGI Video Synthesis

Abstract

A key challenge in photoplethysmography imaging (PPGI) is the limited availability of heterogeneous data. Synthetic data offers a potential solution, but accurately reproducing realistic skin color variations in 3D render engine pipelines depends on multiple parameters, particularly subsurface color and radius. This work proposes a method to analyze the influence of these parameters using spatial variations in amplitude and phase from Pulse 3D Face, enabling more accurate synthetic data generation. A Bayesian optimizer is applied to fit subsurface color and radius scales to match real amplitudes. The results show that subsurface color changes have a greater influence on the resulting amplitudes, as improper scaling can increase the error by up to a factor of 6.6, while poor scaling in subsurface radius increased the error approximately 70 %.