FreqPhys: Repurposing Implicit Physiological Frequency Prior for Robust Remote Photoplethysmography

Wei Qian; Dan Guo; Jinxing Zhou; Bochao Zou; Zitong Yu; Meng Wang

cs.CVApr 1, 2026

FreqPhys: Repurposing Implicit Physiological Frequency Prior for Robust Remote Photoplethysmography

Authors: Wei Qian, Dan Guo, Jinxing Zhou, Bochao Zou, Zitong Yu, Meng Wang

Organizations: Hefei University of Technology · Hefei Comprehensive National Science Center · MBZUAI · University of Science and Technology Beijing ·

Abstract

Remote photoplethysmography (rPPG) enables contactless physiological monitoring by capturing subtle skin-color variations from facial videos. However, most existing methods predominantly rely on time-domain modeling, making them vulnerable to motion artifacts and illumination fluctuations, where weak physiological clues are easily overwhelmed by noise. To address these challenges, we propose FreqPhys, a frequency-guided rPPG framework that explicitly leverages physiological frequency priors for robust signal recovery. Specifically, FreqPhys first applies a Physiological Bandpass Filtering module to suppress out-of-band interference, and then performs Physiological Spectrum Modulation together with adaptive spectral selection to emphasize pulse-related frequency components while suppress residual in-band noise. A Cross-domain Representation Learning module further fuses these spectral priors with deep time-domain features to capture informative spatial--temporal dependencies. Finally, a frequency-aware conditional diffusion process progressively reconstructs high-fidelity rPPG signals. Extensive experiments on six benchmarks demonstrate that FreqPhys yields significant improvements over state-of-the-art approaches, particularly under challenging motion conditions. It highlights the importance of explicitly modeling physiological frequency priors. The source code will be released.

Figures & tables

Figure 1: Visualization of the differences between ground-truth and raw rPPG signals in time and frequency domains. (a) Ground-truth rPPG signal, whose spectrum exhibits clear physiological priors: ( i ) Physiological Band Constraint , where the spectral energy is concentrated within the physiological bandwidth of [0.66, 3.0] Hz corresponding to normal heart rate ranges, and ( ii ) Dominant Peak Property , where a prominent spectral peak within this band reflects the periodic cardiac rhythm, while other in-band components remain comparatively low. The dominant frequency peak (marked by ★ ) corresponds to heart rate and is converted to beats per minute by multiplying the frequency by 60. (b)(c) Raw rPPG signals extracted from facial videos under stable and motion conditions, computed by averaging green-channel pixel intensities over time [ 39 ] . In the time domain, physiological signals are heavily entangled with noise. In the frequency domain, noise manifests as both out-of-band interference and residual in-band components, predominantly concentrated at lower frequencies. While the stable scenario preserves a relatively distinct spectral peak, motion artifacts disperse the in-band energy distribution, making reliable denoising considerably more challenging.

Figure 2: The pipeline of proposed FreqPhys . Given a facial video, we first construct MSTmap $\mathbf{X}$ as the temporal condition and generate the frequency condition $\mathbf{C}^{\mathbf{P}}$ by applying the PBF. During training, we initially generate noise rPPG $\mathbf{Y}_{k}$ by adding Gaussian noise to Ground Truth rPPG $\mathbf{Y}_{0}$ for the $k$ -th step. Then, we input $\mathbf{Y}_{k}$ , $\mathbf{X}$ , $k$ , and $\mathbf{C}^{\mathbf{P}}$ into the Denoising Network . Specifically, the frequency condition $\mathbf{C}^{\mathbf{P}}$ is fed into the Physiological Frequency Denoiser module to enhance physiological spectral clues through three key steps: ( i ) PBF removes out-of-band noise based on the physiological frequency bandwidth [0.66,3.0] Hz; ( ii ) PSM emphasizes valid physiological harmonics by modeling interactions between real and imaginary components; ( iii ) ASS dynamically suppresses in-band noise using data-driven energy thresholds. Next, with Cross-domain Representation Learning , our FreqPhys includes frequency-domain denoised information into space and time dependencies modeling to estimate the high-fidelity rPPG signal. During inference, the initial rPPG $\mathbf{Y}_{K}$ is randomly sampled from Gaussian noise, with frequency condition and denoising network processes mirroring those used in training.

Figure 3: The details of physiological spectrum modulation module .

Figure 4: Architecture comparison with existing diffusion methods .

Traditional rPPG Methods
		UBFC-rPPG			PURE			BUAA			MMPD			VIPL-HR			MR-NIRP-Car
Method	Venue	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$
GREEN [ 38 ]	Opt Express’08	7.50	14.41	0.62	7.23	17.05	0.69	6.89	10.39	0.60	21.68	27.69	-0.01	-	-	-	11.88	14.46	0.17
ICA [ 27 ]	Opt Express’10	5.17	11.76	0.65	3.76	12.60	0.85	-	-	-	18.60	24.30	0.01	-	-	-	10.87	13.62	0.02
CHROM [ 9 ]	TBE’13	2.37	4.91	0.89	2.07	2.50	0.99	-	-	-	13.66	18.76	0.08	11.40	16.90	0.28	7.16	9.09	0.35
Deep Learning-based rPPG Methods

Table 1: Intra-dataset HR estimation results. Models are trained and tested on the UBFC-rPPG, PURE, VIPL-HR, and MMPD datasets. Bold : best results.

Traditional rPPG Methods
		LF (n.u.)			HF (n.u)			LF/HF			RF (Hz)
Method	Venue	SD $\downarrow$	RMSE $\downarrow$	$r\uparrow$	SD $\downarrow$	RMSE $\downarrow$	$r\uparrow$	SD $\downarrow$	RMSE $\downarrow$	$r\uparrow$	SD $\downarrow$	RMSE $\downarrow$	$r\uparrow$
GREEN [ 38 ]	Opt Express’08	0.186	0.186	0.280	0.186	0.186	0.280	0.361	0.365	0.492	0.087	0.086	0.111
ICA [ 27 ]	Opt Express’10	0.243	0.240	0.159	0.243	0.240	0.159	0.655	0.645	0.226	0.086	0.089	0.102
POS [ 39 ]	TBE’16	0.171	0.169	0.479	0.171	0.169	0.479	0.405	0.399	0.518	0.109	0.107	0.087
Deep Learning-based rPPG Methods

Table 2: HRV and RF estimation results. Models are trained and tested on the UBFC-rPPG dataset. LF, HF, and RF represent low frequency, high frequency, and respiration frequency, respectively. “n.u.” denotes normalized units.

Traditional rPPG Methods
		PURE $\rightarrow$ UBFC-rPPG			UBFC-rPPG $\rightarrow$ PURE			PURE $\rightarrow$ MMPD			UBFC-rPPG $\rightarrow$ MMPD
Method	Venue	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$
GREEN [ 38 ]	Opt Express’08	19.73	21.00	0.37	10.09	23.85	0.34	21.68	27.69	-0.01	21.68	27.69	-0.01
ICA [ 27 ]	Opt Express’10	16.00	25.65	0.44	4.77	16.07	0.72	18.06	24.30	0.01	18.06	24.30	0.01
CHROM [ 9 ]	TBE’13	4.06	8.83	0.89	5.77	14.93	0.81	13.66	18.76	0.08	13.66	18.76	0.08
Deep Learning-based rPPG Methods

Table 3: Cross-dataset HR estimation results. Models are trained on PURE/UBFC-rPPG, and tested on UBFC-rPPG/PURE/MMPD.

Table 8

Method	Parameters	Flops	Throughput	Inference	Memory
Method	(M) $\downarrow$	(G) $\downarrow$	(Kps) $\uparrow$	time (ms) $\downarrow$	(M) $\downarrow$
DeepPhys [ 7 ]	1.98	111.67	28.89	34.61	10638
PhysNet [ 44 ]	0.77	65.74	68.73	14.55	3750
TS-CAN [ 18 ]	1.98	111.67	26.23	38.13	11834
PhysFormer [ 46 ]	7.38	47.44	50.79	19.69	6480
EfficientPhys [ 19 ]	1.91	56.06	41.36	24.18	7814

Table 6: Computational cost . Best results are marked in bold and the second best in underline.

Figure 5: Time and frequency domain visualizations of rPPG signal predictions on the VIPL dataset under head motion scenario. In the frequency-domain plots, the purple dashed box indicates the physiological signal bandwidth ranging from 0.66 to 3.0 Hz, corresponding to typical human cardiac frequencies. ${\color[rgb]{1,0,0}\star}$ and ${\color[rgb]{0.3086,0.6797,0.3555}\bullet}$ denote the spectral peaks of the ground-truth and predicted heart rates, respectively.

Figure 6: HR estimation results on VIPL-HR under different scenarios and cameras.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Description
$\mathbf{V}$	Facial video sequence.
$T$	Temporal dimension of the input video sequence.
$H$	Spatial height of the input video.
$W$	Spatial width of the input video.
$\mathbf{X}$	MSTmap used as model input.
$N$	Facial regions of interest (ROIs).

Appendix

Table 7: A list of key notations and their interpretation in this work.

Figure 7: Implementation details of the stacked denoising network in our FreqPhys .

Figure 8: Typical samples from six different remote physiological measurement datasets.

Traditional rPPG Methods
		U $\rightarrow$ C			M $\rightarrow$ C			V $\rightarrow$ C			Average
Method	Venue	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$
GREEN [ 38 ]	Opt Express’08	11.01	14.60	0.19	11.01	14.60	0.19	11.01	14.60	0.19	11.01	14.60	0.19
ICA [ 27 ]	Opt Express’10	10.07	13.89	0.28	10.07	13.89	0.28	10.07	13.89	0.28	10.07	13.89	0.28
CHROM [ 9 ]	TBE’13	9.70	13.71	0.28	9.70	13.71	0.28	9.70	13.71	0.28	9.70	13.71	0.28
POS [ 39 ]	TBE’16	9.32	13.56	0.21	9.32	13.56	0.21	9.32	13.56	0.21	9.32	13.56	0.21

Appendix

Table 8: Cross-domain generalization evaluation results on the MR-NIRP-Car dataset. U=UBFC-rPPG, M=MMPD, V=VIPL-HR, C=MR-NIRP-Car. Bold : best results.

Traditional rPPG Methods
		P+U $\rightarrow$ B			P+M $\rightarrow$ B			U+M $\rightarrow$ B			Average
Method	Venue	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$
GREEN [ 38 ]	Opt Express’08	6.89	10.39	0.60	6.89	10.39	0.60	6.89	10.39	0.60	6.89	10.39	0.60
CHROM [ 9 ]	TBE’13	6.09	8.29	0.51	6.09	8.29	0.51	6.09	8.29	0.51	6.09	8.29	0.51
POS [ 39 ]	TBE’16	5.04	7.12	0.63	5.04	7.12	0.63	5.04	7.12	0.63	5.04	7.12	0.63
Deep Learning-based rPPG Methods

Appendix

Table 9: Cross-domain generalization evaluation results on the BUAA dataset, with dual-source training and single-target testing. U=UBFC-rPPG, P=PURE, B=BUAA, and M=MMPD. Bold : best results.

Traditional rPPG Methods
		P+U $\rightarrow$ M			P+B $\rightarrow$ M			U+B $\rightarrow$ M			Average
Method	Venue	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$
GREEN [ 38 ]	Opt Express’08	21.68	27.69	-0.01	21.68	27.69	-0.01	21.68	27.69	-0.01	21.68	27.69	-0.01
CHROM [ 9 ]	TBE’13	13.66	18.76	0.08	13.66	18.76	0.08	13.66	18.76	0.08	13.66	18.76	0.08
POS [ 39 ]	TBE’16	12.36	17.71	0.18	12.36	17.71	0.18	12.36	17.71	0.18	12.36	17.71	0.18
Deep Learning-based rPPG Methods

Appendix

Table 10: Cross-domain generalization evaluation results on the MMPD dataset, with dual-source training and single-target testing. U=UBFC-rPPG, P=PURE, B=BUAA, and M=MMPD. Bold : best results.

Traditional rPPG Methods
		P+B+M $\rightarrow$ U			U+B+M $\rightarrow$ P			P+U+M $\rightarrow$ B			P+U+B $\rightarrow$ M
Method	Venue	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$
GREEN [ 38 ]	Opt Express’08	19.73	31.00	0.37	10.09	23.85	0.34	6.89	10.39	0.60	21.68	27.69	-0.01
CHROM [ 9 ]	TBE’13	7.23	8.92	0.51	9.79	12.76	0.37	6.09	8.29	0.51	13.66	18.76	0.08
POS [ 39 ]	TBE’16	7.35	8.04	0.49	9.82	13.44	0.34	5.04	7.12	0.63	12.36	17.71	0.18
Deep Learning-based rPPG Methods

Appendix

Table 11: Cross-domain generalization evaluation results on the UBFC-rPPG/PURE/BUAA/MMPD datasets, with three-source training and single-target testing. U=UBFC-rPPG, P=PURE, B=BUAA, and M=MMPD. Bold : best results.

Method	VIPL-HR			MR-NIRP-Car
Method	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$
w/o Diffusion	3.98	6.64	0.82	5.88	6.51	0.52
with Diffusion (Ours)	3.79	6.34	0.86	5.75	6.08	0.54

Appendix

Table 12: Impact of diffusion model.

Method	[40, 60)	[60, 100)	[100, 180]
PhysDiff [ 29 ]	8.73	4.57	13.40
FreqPhys (Ours)	6.83	4.25	10.73

Appendix

Table 13: Comparison of different heart rate ranges. We provide the RMSE

\downarrow

results on the VIPL-HR dataset.

Method	MAE $\downarrow$	RMSE $\downarrow$	$r\uparrow$
1	7.42	10.60	0.66
2	4.84	7.82	0.81
3	3.90	6.52	0.86
4	3.79	6.34	0.86
5	3.97	6.60	0.85
6	5.19	7.74	0.82

Appendix

Table 14: Ablation study on the number of cross-attention layers in the Cross-domain Representation Learning module on the VIPL-HR dataset.

Remote Photoplethysmography

Remote photoplethysmography (rPPG) estimates pulse and other physiological signals from video without contact sensors.

See all 46 papers

Jun 22, 2026

PhysFlow: Frequency Decoupled with Dual-Field Rectified Flow for Remote Photoplethysmography

Remote Photoplethysmography (rPPG) enables contactless pulse estimation from facial videos, serving as a vital tool for health monitoring. However, current deep learning methods often struggle under complex disturbances, particularly varying illumination, facial expressions, and unconstrained head movements. In such scenarios, subtle physiological signals are easily dominated by external interference, making the recovered rPPG waveform unstable and unreliable. One important reason is that most existing methods directly model the rPPG signal in a unified manner, where different signal components are coupled during reconstruction. This makes it difficult to preserve weak pulse-related variations when strong disturbance-induced changes are present. To address this challenge, we propose PhysFlow, a frequency-decoupled dual-field rectified flow framework tailored for robust rPPG estimation. Specifically, the ground-truth rPPG signal is decomposed into trend and amplitude components, which are used as separate supervisory targets. Based on the extracted facial features, PhysFlow learns two component-specific conditional velocity fields to model the two components separately. This design reduces mutual interference between different components and improves the robustness of rPPG reconstruction under complex disturbances. Moreover, the rectified flow formulation enables efficient waveform reconstruction with only a few ordinary differential equation (ODE) integration steps. Extensive experiments on multiple benchmark datasets demonstrate that PhysFlow outperforms state-of-the-art methods in both heart-rate estimation and rPPG waveform reconstruction across diverse challenging scenarios.

Zixu Li, jianjun Qian, Hang Shao +2

PCA Lab, Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing 210094, China · College of Computer Science and Technology, Qingdao University, Qingdao 266071, China

Remote Photoplethysmography

Ask PDF

Mar 20, 2026

PhysNeXt: Next-Generation Dual-Branch Structured Attention Fusion Network for Remote Photoplethysmography Measurement

Remote photoplethysmography (rPPG) enables contactless measurement of heart rate and other vital signs by analyzing subtle color variations in facial skin induced by cardiac pulsation. Current rPPG methods are mainly based on either end-to-end modeling from raw videos or intermediate spatial-temporal map (STMap) representations. The former preserves complete spatiotemporal information and can capture subtle heartbeat-related signals, but it also introduces substantial noise from motion artifacts and illumination variations. The latter stacks the temporal color changes of multiple facial regions of interest into compact two-dimensional representations, significantly reducing data volume and computational complexity, although some high-frequency details may be lost. To effectively integrate the mutual strengths, we propose PhysNeXt, a dual-input deep learning framework that jointly exploits video frames and STMap representations. By incorporating a spatio-temporal difference modeling unit, a cross-modal interaction module, and a structured attention-based decoder, PhysNeXt collaboratively enhances the robustness of pulse signal extraction. Experimental results demonstrate that PhysNeXt achieves more stable and fine-grained rPPG signal recovery under challenging conditions, validating the effectiveness of joint modeling of video and STMap representations. The codes will be released.

Junzhe Cao, Bo Zhao, Zhiyi Niu +5

Harbin Institute of Technology, Shenzhen · Great Bay University · Hefei University of Technology +1

Remote Photoplethysmography Heart Rate Estimation

Ask PDF

Jun 22, 2026

StreamPPG: Low-Latency rPPG Estimation via Consistent Privileged Learning

Remote photoplethysmography (rPPG) estimates the blood volume pulse (BVP) signal from facial videos, enabling contact-free health monitoring. Conventional clip-wise approaches, which use video clips as input, require capturing over one hundred frames before inference, thus introducing several seconds of delay and hindering real-time use. Meanwhile, frame-wise approaches struggle to capture long-range temporal and periodic features of physiological rhythms, and therefore lead to reduced estimation accuracy. To overcome these issues, we propose StreamPPG, a unified architecture that enables low-latency frame-wise physiological signal estimation while achieving competitive accuracy compared with clip-wise approaches. StreamPPG is trained under a consistent privileged learning (CPL) strategy, which leverages ground-truth rPPG signals as privileged information to enhance the model's representation capability. Extensive experiments demonstrate that StreamPPG achieves state-of-the-art accuracy across multiple datasets while maintaining real-time throughput on edge devices.

Yiming Li, Yihan Yang, Yuguang Chu +4

ARXIV

Remote Photoplethysmography Photoplethysmography

Ask PDF

Abstract

Remote Photoplethysmography

Explore similar work

PhysFlow: Frequency Decoupled with Dual-Field Rectified Flow for Remote Photoplethysmography

PhysNeXt: Next-Generation Dual-Branch Structured Attention Fusion Network for Remote Photoplethysmography Measurement

StreamPPG: Low-Latency rPPG Estimation via Consistent Privileged Learning