PulsePath: Two-Stage Generative Waveform Refinement for Robust Remote Photoplethysmography Estimation

Remote photoplethysmography (rPPG) estimates pulse signals from facial videos, but motion, illumination variation, exposure control, and compression can distort the waveform used for heart-rate (HR) estimation. The Plane-Orthogonal-to-Skin (POS) waveform provides an observed physiological source that often preserves recording-specific periodic structure. Existing methods commonly regress the complete waveform directly or reconstruct it from a stochastic state without using this source as the starting point. We propose PulsePath, a two-stage framework that transports POS toward paired-contact photoplethysmography (PPG) and models only the residual remaining after deterministic inference. Stage A performs conditional Flow Matching (FM) from POS using a six-channel multiscale spatiotemporal map and a 25-channel recording-level frequency representation. Stage B fixes the deterministic FM estimate and reconstructs its remaining error with variance-preserving residual diffusion and velocity prediction. In intra-dataset testing, PulsePath achieves mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation values of 0.46/0.94/0.99 on PURE and 0.29/0.80/0.99 on UBFC-rPPG. In cross-dataset testing, PulsePath achieves an MAE/RMSE of 0.23/0.63 when trained on UBFC-rPPG and tested on PURE, and 1.13/2.49 when trained on PURE and tested on UBFC-rPPG. The additional analyses show protocol-dependent component effects and training-run variability; the reported HR results do not establish waveform fidelity.