Data-Augmentation for Deep Learning Based Remote Photoplethysmography Methods

Acquisition of remote photoplethysmography (rPPG) signals using deep learning-based methods has become very important for the measurement of heart rate (HR). These methods are known to require a large amount of data during training, so data augmentation is often used. In this paper we propose a methodology for data augmentation to be used as a pre-training step. We tested our proposed method using transfer-learning in three public databases, and we demonstrate that it helps the neural network to learn the main features of the videos. We improved the Mean Absolute Error by a factor of 3 using the small dataset UBFC-rPPG, by a factor of 2.5 in the medium-size dataset COHFACE and by a factor of 1.3 in the large dataset VIPL-HR.