Title: Template Collapse and Information-Theoretic Limits in Camera rPPG Pulse Morphology Restoration

Abstract

Objective: Remote photoplethysmography (rPPG) using consumer-grade face cameras facilitates passive cardiovascular monitoring. However, it remains unclear whether the specific morphology of single-cycle waveforms, which encodes arterial stiffness biomarkers, can be successfully recovered from these measurements.

Methods: This study assessed 16 distinct architectures belonging to six different families, utilizing data from 153 subjects drawn across three separate datasets. To differentiate between genuine subject-specific recovery and template collapse, we introduced the cross-subject Pearson correlation coefficient (r) as a diagnostic metric.

Results: None of the evaluated architectures were able to recover subject-specific morphology, with cross-subject r values ranging from 0.773 to 0.9999, compared to a ground-truth ceiling of 0.601. The Supervised Contrastive (SupCon) model converged at log N = 4.844, providing the most robust empirical evidence to date that the encoder families tested cannot extract discriminative morphological structures from single-cycle rPPG signals. While the Variational Autoencoder (VAE) decoder successfully restored population-level harmonic content that was absent in the original rPPG input (yielding an H2/H1 ratio of 0.310 in the output versus 0.275 in the input), this capability allowed for zero-shot generalization to the UBFC dataset (r = +0.708). A directional hallucination gap (p = 0.150) indicated a partial reading of the signal. Furthermore, anti-collapse objectives proved ineffective when the input data lacked discriminative structure.

Significance: The findings indicate that consumer cameras are incapable of encoding individual arterial morphology. Consequently, cross-subject r serves as an essential diagnostic tool for identifying collapse in waveform reconstruction benchmarks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC