Self-Supervised Augmented Vision Transformers for Remote Physiological Measurement

Video based heart rate (HR) measurement conducting remote photoplethysmography (rPPG) exhibits immense prospects and achieves creditable results in constrained scenarios. (1) Background: Non-contact facial video-based HR estimation is a constantly evolving direction, in which deep learning models achieved state-of-the-art performance. However, these methods usually have higher computational burden, which are difficult to deploy on mobile or edge devices directly; (2) Methods: To address this problem, we propose a lightweight model for HR estimation based on rPPG. This model uses vision transformer to serves as the backbone, and the contraction factor is adopted to reduce the model complexity. (3) Results: The proposed model is tested on the challenging VIPL-HR-V2 dataset. The results show that the proposed model can greatly reduce the model complexity while attaining reasonable accuracy.