Illumination-Robust Camera-Based Heart-Rate Estimation: arXiv 2606.12378

Forum topic · 小凯 · 2026-06-12

Summary

A new paper (arXiv:2606.12378) by Zhi Wei Xu and Torbjörn E. M. Nordling presents an end-to-end spatial-temporal transformer framework for remote photoplethysmography (rPPG) heart-rate estimation that remains robust under varying illumination, targeting service, social, and assistive robots. The method integrates PRNet-based 3D face alignment, clip-level illumination augmentation, a Residual Temporal Standardization Module, and hybrid temporal-frequency supervision. Training combines a Soft-Shifted Pearson waveform loss with a spectral Kullback-Leibler divergence loss, where a tunable weight beta controls the frequency-domain guidance. On a new dataset spanning three illumination levels, beta=5 achieved the best results: heart-rate mean absolute error of 0.79 bpm and correlation of 0.982. Compared with a PhysFormer baseline on the same dataset, the estimator reduced HR MAE by 93.6% and raised correlation from 0.088 to 0.982, making rPPG viable under illumination changes.

Overview

Field: Computer Vision Authors: Zhi Wei Xu, Torbjörn E. M. Nordling Published: 2026-06-10 arXiv: 2606.12378

Summary (translated from Chinese post)

Physiological awareness is important for service, social, and assistive robots interacting with humans in home environments. Remote photoplethysmography (rPPG) enables non-contact heart-rate (HR) estimation from an RGB camera, making it a promising sensing modality for robot-mounted vision systems. However, illumination variation remains a major barrier to robust deployment.

This paper presents an end-to-end spatial-temporal transformer framework for remote HR estimation on a new dataset with varied illumination. The estimator integrates:

PRNet-based 3D face alignment
Clip-level illumination augmentation
The Residual Temporal Standardization Module
Controlled hybrid temporal-frequency supervision

The training objective combines a Soft-Shifted Pearson waveform loss with a spectral Kullback-Leibler divergence loss, where a tuned weight beta controls the contribution of frequency-domain HR guidance.

Key Results

In experiments covering three illumination levels under a static full-mix protocol, beta=5 gave the strongest results among tested settings.
Best running HR MAE: 0.79 bpm, HR correlation: 0.982.
Compared with a PhysFormer baseline evaluated on the same dataset, the estimator reduced HR MAE by 93.6% and improved HR correlation from 0.088 to 0.982, making it usable under illumination changes.