Remote Physiological Monitoring: Domain Shifts and Their Effects on Model Behavior

Remote physiological monitoring is a category of computer vision problems in which physiological features are inferred from video data. This dissertation focuses on the challenging task of remote photoplethysmography (rPPG), a technique for inferring a subject's heart rate and its blood volume pulse waveform from video. In particular, rPPG leverages the minute color fluctuations on the subject's skin due to the light absorption of hemoglobin to arrive at the blood volume pulse waveform. Using rPPG as our focal problem, we explore generalization under domain shifts by releasing unique and challenging datasets, developing model training regimes that promote generalization for the rPPG problem, proposing techniques for explaining why models behave as they do even under domain shifted applications, and providing tools that could detect domain shifted data in an in-the-wild application.