Learning Motion-Robust Remote Photoplethysmography through Arbitrary Resolution Videos

Remote photoplethysmography (rPPG) enables non-contact heart rate (HR) estimation from facial videos which gives significant convenience compared with traditional contact-based measurements. In the real-world long-term health monitoring scenario, distance of participants and their head movements usually vary by time, resulting in inaccurate rPPG measurement due to varying face resolution complex motion artifacts. Different previous models designed for a constant between camera participants, this paper, we propose two plug-and-play blocks (i.e., physiological signal feature extraction block (PFE) temporal alignment (TFA)) alleviate degradation changing motion. On one side, guided representative-area information, PFE adaptively encodes arbitrary frames fixed-resolution structure features. other leveraging estimated optical flow, TFA is able counteract confusion caused movement thus benefit motion-robust recovery. Besides, also train model cross-resolution constraint using two-stream dual-resolution framework, further helps learn resolution-robust Extensive experiments on three benchmark datasets (UBFC-rPPG, COHFACE PURE) demonstrate superior performance proposed method. One highlight that TFA, off-the-shelf spatio-temporal can predict more robust signals under both severe scenarios. The codes are available at https://github.com/LJWGIT/Arbitrary_Resolution_rPPG.