candidate-aware rPPG HR with an auditable release/review output contract (Research Square 2026)
Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.
Instead of trusting one rPPG estimator, VitalsSight collects pulse candidates from many routes: multi-ROI classical methods, learned networks and transformers, plus optional temporal-correction candidates. A relation-aware selector then picks one. The output is split into three parts: the selected HR, an evidence packet describing the candidate ambiguity, and a proposed release/review state.
multi-route candidate construction (multi-ROI classical, learned and transformer rPPG routes, optional temporal-correction candidates) → physiology-aware evidence features → relation-aware set-attention / graph selector (harmonic penalty, KL soft target) → selected HR + evidence packet + proposed release/review state (fixed-coverage gate)
Camera-based contactless heart rate monitoring offers significant potential for clinical care and longitudinal research, yet estimation often fails in challenging windows containing competing pulse candidates, aliases, or motion artifacts. To overcome this deployment bottleneck, we present VitalsSight, a candidate-aware supervised framework that retains pulse candidates across diverse signal routes prior to selection, thereby preserving critical physiological evidence. In internal subject-disjoint evaluations on the University of Bourgogne Franche-Comté remote photoplethysmography (UBFC-rPPG) benchmark dataset (42 participants, 439 windows), VitalsSight achieved a primary window-level mean absolute error of 1.646 ± 0.051 beats per minute, with consistent error reductions across stress datasets including MCD-rPPG and rPPG-10. Crucially, when encountering real-world deployment boundaries—such as severe low light and source shifts—VitalsSight outputs an evidence packet capturing candidate ambiguity to support a transparent release/review contract. Rather than masking errors through forced automated outputs, this auditable mechanism establishes a safer and more pragmatic foundation for the clinical translation of digital vital sign monitoring systems. Code available at https://github.com/ShikiRyo1/VitalsSight-candidate-aware-rPPG.
Verbatim fromthe publisher's DOI record· Publisher licence: https://creativecommons.org/licenses/by/4.0/
Datasets linked in the hub; pretraining vs evaluation roles follow the paper.
Full photo credits, sources and licences are on each device page.
UBFC-Physmultimodal social-stress (wrist BVP+EDA + face video)Evaluation56 subjects · PPG · rPPG · EDAVitalsSight. (2026). Research Square preprint. https://doi.org/10.21203/rs.3.rs-10673444/v1
@misc{vitalssight2026vitalssight,
title = {{VitalsSight}},
year = {2026},
howpublished = {Research Square preprint},
doi = {10.21203/rs.3.rs-10673444/v1}
}Record updated