Researchers propose intervention-based self-supervised learning paradigm to improve remote photoplethysmography signal detection in video

A new framework called Interv-rPPG treats the latent rPPG (remote photoplethysmography) signal as an underlying physical source and pixel chrominance variations as its visual manifestation, shifting from passive correlation learning to active intervention on video based on an rPPG hypothesis.

The framework uses PhysMambaFormer (an rPPG extractor) to hypothesize the rPPG signal and a Controllable Physiological Signal Editor to conduct precise chrominance-domain interventions, validating physical realism through 'Falsifiability via Nulling' and 'Axiomatic Equivariance'.

The method improves both in-domain and cross-domain performance on datasets such as VIPL-HR and MMPD, surpasses supervised baselines in complex cross-dataset settings, and effectively resists motion and illumination artifacts according to diagnostic analysis of nuisance sensitivity.