Remote Heart Rate Estimation from Facial Videos with Balanced Contrastive Learning

Dong-Hyuk Lee, Dae Hyeon Kim, Young-Seok Choi

Abstract: Measuring remote photoplethysmography (rPPG), a contactless facial video-based PPG estimation, requires a large amount of labeled data via supervised methods, leading to significantly increased labor and costs. Existing rPPG studies based on self-supervised methods have utilized temporal and spatial similarities between rPPG data, which help reduce manual labeling costs. Here, we propose a novel self-supervised rPPG estimation method with contrastive learning. The proposed method not only leverages temporal and spatial similarities as representations but also maintains a well-contained representation. A well-contained representation reduces the computational cost associated with calculating negative loss. We use a 3D-CNN model to capture well-contained representation and multiple rPPG signals, which perform on different spaces but are temporally similar. We show that the proposed method outperforms current rPPG methods in terms of heart rate estimation accuracy on two public datasets, i.e., PURE and UBFC.