In pre-clinical health monitoring, estimating physiological signals from video is a low-cost and convenient tool. Remote photoplethysmography (rPPG) involves placing a camera in a remote area to
estimate a person’s heart rate or Blood Volume Pulse (BVP). In this paper, we propose an attention based deep architecture for rPPG estimation that assimilate temporal relationship across a sequence
of frames while focusing on the relevant features and regions by exploiting the inter-pixel relationship of feature maps. Also, we design a dynamic supervision strategy using frequency and time domain
losses to mitigate overfitting for efficient estimation of rPPG signals. The proposed method was evaluated on two publicly available rPPG datasets (UBFC-rPPG and PURE). The findings of this study
demonstrate that promising results can be achieved by enforcing an adequate balance between time-frequency supervision.