Title: FCUS-rPPG: An Efficient Unsupervised Approach to Remote Photoplethysmography Using Gradient Oscillation Mitigation

Abstract

Remote photoplethysmography (rPPG) facilitates the contactless acquisition of blood volume pulse (BVP) signals through the use of standard cameras. While recent unsupervised rPPG techniques have emerged to learn BVP representations without the need for labeled physiological data, their training processes are frequently plagued by unstable and noisy gradients. This instability often leads to sluggish convergence and poor generalization across different domains. To address these challenges, we introduce FCUS-rPPG, a novel unsupervised rPPG framework characterized by rapid convergence and robust generalization abilities.

Drawing from the insight that BVP representations possess both multi-spectral covariation and a low-dimensional manifold structure, we engineered a spectrally shared backbone. This architectural choice aids in disentangling BVP features while simultaneously boosting optimization efficiency. Furthermore, we established a comprehensive optimization framework that operates across three distinct levels: the gradient, the loss landscape, and feature representation. This holistic approach aims to improve both the stability of convergence and the overall generalization performance.

Our method incorporates three key innovations: 1. Post-verification masking: This mechanism eliminates misleading gradients by leveraging the weak-amplitude physiological prior inherent in BVP signals. 2. Perturbation-based loss landscape smoothing: This strategy guides the optimization process toward flatter minima, which are known to be more generalizable. 3. Noise-aware null-space regularization: This technique restricts feature updates to the orthogonal complement of the noise subspace, effectively reducing representation drift caused by noise.

We evaluated FCUS-rPPG through extensive experiments across five datasets. The results indicate that FCUS-rPPG achieves optimal performance after just a single training epoch, a stark contrast to existing methods that typically demand tens or even hundreds of epochs. Moreover, FCUS-rPPG consistently secured state-of-the-art (SOTA) results in cross-dataset evaluations. This research offers a highly efficient and resilient solution for the practical implementation of unsupervised rPPG. The source code for this project will be made publicly available at https://github.com/JiaJieLee/FCUS-rPPG.

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