KDPhys: An Attention Guided 3D to 2D Knowledge Distillation for Real-time Video-Based Physiological Measurement

TL;DR

KDPhys introduces an attention-guided knowledge distillation framework that effectively extracts remote photoplethysmography signals from facial videos, significantly reducing model complexity and improving real-time performance for non-contact physiological monitoring.

Contribution

This work is the first to apply knowledge distillation to the rPPG domain, distilling 3D CNN features into a lightweight 2D CNN with a novel DILATE loss for enhanced signal quality.

Findings

Model uses half the parameters of existing methods.

Achieves 56.67% faster inference speed.

Reduces MAE by 18.15% to 1.78 bpm.

Abstract

Camera-based physiological monitoring, such as remote photoplethysmography (rPPG), captures subtle variations in skin optical properties caused by pulsatile blood volume changes using standard digital camera sensors. The demand for real-time, non-contact physiological measurement has increased significantly, particularly during the SARS-CoV-2 pandemic, to support telehealth and remote health monitoring applications. In this work, we propose an attention-based knowledge distillation (KD) framework, termed KDPhys, for extracting rPPG signals from facial video sequences. The proposed method distills global temporal representations from a 3D convolutional neural network (CNN) teacher model to a lightweight 2D CNN student model through effective 3D-to-2D feature distillation. To the best of our knowledge, this is the first application of knowledge distillation in the rPPG domain. Furthermore, we introduce a Distortion Loss incorporating Shape and Time (DILATE), which jointly accounts for both morphological and temporal characteristics of rPPG signals. Extensive qualitative and quantitative evaluations are conducted on three benchmark datasets. The proposed model achieves a significant reduction in computational complexity, using only half the parameters of existing methods while operating 56.67% faster. With just 0.23M parameters, it achieves an 18.15% reduction in Mean Absolute Error (MAE) compared to state-of-the-art approaches, attaining an average MAE of 1.78 bpm across all datasets. Additional experiments under diverse environmental conditions and activity scenarios further demonstrate the robustness and adaptability of the proposed approach.