Abstract
Remote photoplethysmography (rPPG) is a contactless method for heart rate (HR) estimation from face videos. In this paper, we propose to estimate rPPG signals directly from input video sequences in an end-to-end manner. We propose a novel Siamese-rPPG network to simultaneously learn the heterogeneous and homogeneous features from two facial regions. Furthermore, to analyze the temporal periodicity of rPPG signals, we construct the network with 3D CNNs and jointly train the two-branch model under the negative Pearson loss function. Experimental results on three benchmark datasets: COHFACE, UBFC, and PURE, show that our method significantly outperforms existing methods with a large margin.