LCOMS Lab's Approach to the Vision for Vitals (V4V) Challenge

Yassine Ouzar, Djamaleddine Djeldjli, Frédéric Bousefsaf, Choubeila Maaoui; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 2750-2754

We present in this paper the LCOMS Lab's approach to the 1st Vision For Vitals (V4V) Challenge organized within ICCV2021. The V4V challenge was focused on computer vision methods for vitals signs measurement from facial videos, including pulse rate (PR) and respiratory rate. We propose a novel end-to-end architecture based on a deep spatiotemporal network for pulse rate estimation from facial video recordings. Unlike existing methods, we predict the pulse rate value directly without passing by iPPG signal extraction and without incorporating any prior knowledge or additional processing steps. We built our network using 3D Depthwise Separable Convolution layers with residual connections to extract spatial and temporal features simultaneously. This is very suitable for real-time measurement because it requires a reduced number of parameters and a short video fragment. The obtained results seem very satisfactory and promising, especially since the experiments were conducted in challenging dataset collected in uncontrolled conditions.