Self-Supervised Representation Learning Framework for Remote Physiological Measurement Using Spatiotemporal Augmentation Loss

Recent advances in supervised deep learning methods are enabling remote measurements of photoplethysmography-based physiological signals using facial videos. The performance these methods, however, dependent on the availability large labelled data. Contrastive as a self-supervised method has recently achieved state-of-the-art performances representative data features by maximising mutual information between different augmented views. However, existing augmentation techniques for contrastive not designed to learn from videos and often fail when there complicated noise subtle periodic colour/shape variations video frames. To address problems, we present novel spatiotemporal framework signal representation learning, where is lack training Firstly, propose landmark-based spatial that splits face into several informative parts based Shafer’s dichromatic re?ection model characterise skin colour fluctuations. We also formulate sparsity-based temporal exploiting Nyquist–Shannon sampling theorem effectively capture changes modelling features. Furthermore, introduce constrained loss which generates pseudo-labels clips. It used regulate process handle noise. evaluated our 3 public datasets demonstrated superior than other competitive accuracy compared methods. Code available at https://github.com/Dylan-H-Wang/SLF-RPM.