Cross-View PSD Distillation for Viewpoint-Robust Remote Photoplethysmography

Abstract

Remote photoplethysmography (rPPG) enables contactless heart rate measurement from facial videos, but models trained on frontal views suffer significant degradation on side views, limiting real-world deployment. We propose asymmetric cross-view PSD distillation, which transfers frequency-domain knowledge from a reliable frontal view to side views during training. The key insight is that while visual appearance varies across viewpoints, the power spectral density (PSD) of the cardiac signal remains view-invariant. Critically, we apply a stop-gradient operation to the teacher (frontal) view, preventing the PSD loss from corrupting the high-quality frontal branch---symmetric approaches cause catastrophic degradation. On the MCD-rPPG dataset, our method reduces side-view heart rate MAE from 5.24 to 3.99 bpm (24% improvement) while simultaneously improving frontal performance from 2.91 to 2.45 bpm, reducing the frontal-to-side gap by 34%. This enables more robust rPPG deployment in scenarios with varying camera angles.