rPPG-VQA: A Video Quality Assessment Framework for Unsupervised rPPG Training

Tianyang Dai; Ming Chang; Yan Chen; Yang Hu

cs.CVApr 13, 2026

rPPG-VQA: A Video Quality Assessment Framework for Unsupervised rPPG Training

Authors: Tianyang Dai, Ming Chang, Yan Chen, Yang Hu

Organizations: University of Science and Technology of China

Abstract

Unsupervised remote photoplethysmography (rPPG) promises to leverage unlabeled video data, but its potential is hindered by a critical challenge: training on low-quality "in-the-wild" videos severely degrades model performance. An essential step missing here is to assess the suitability of the videos for rPPG model learning before using them for the task. Existing video quality assessment (VQA) methods are mainly designed for human perception and not directly applicable to the above purpose. In this work, we propose rPPG-VQA, a novel framework for assessing video suitability for rPPG. We integrate signal-level and scene-level analyses and design a dual-branch assessment architecture. The signal-level branch evaluates the physiological signal quality of the videos via robust signal-to-noise ratio (SNR) estimation with a multi-method consensus mechanism, and the scene-level branch uses a multimodal large language model (MLLM) to identify interferences like motion and unstable lighting. Furthermore, we propose a two-stage adaptive sampling (TAS) strategy that utilizes the quality score to curate optimal training datasets. Experiments show that by training on large-scale, "in-the-wild" videos filtered by our framework, we can develop unsupervised rPPG models that achieve a substantial improvement in accuracy on standard benchmarks. Our code is available at https://github.com/Tianyang-Dai/rPPG-VQA.

Figures & tables

Figure 1: Dataset categorization and performance comparison for unsupervised rPPG. (a) Classification of datasets into labeled rPPG benchmarks and unlabeled “in-the-wild” corpora. (b) HR estimation errors on the PURE test set using different sampling criteria. “All” and “Random” denote the full and randomly sampled combined dataset (CAS(ME) 2 and Celeb-DF (v2)), respectively.

Figure 2: An overview of our proposed rPPG-VQA framework. The method employs a dual-branch architecture to assess video quality. The signal-level noise perception branch evaluates physiological signal integrity by fusing SNR estimates from multiple rPPG algorithms based on their consensus. The scene-level noise perception branch uses an MLLM to score interferences like motion and illumination. The two assessments are combined into a unified quality score, which guides a two-stage adaptive sampling strategy to construct the final target training set.

Dimension	Score
Head Movement Noise	$\left\{{0,1,2,3}\right\}$
Illumination Noise	$\left\{{0,1,2,3}\right\}$
Skin Noise	$\left\{{0,1,2}\right\}$
Camera Noise	${\gamma_{i}}\cdot\left\{{0,1,2}\right\}$

Table 1: Scene noise dimensions and scoring scales for MLLM-based assessment. Higher scores denote better quality.

Training Set	Sampling	Quality Criterion	Gideon21			Contrast-Phys			SiNC
Training Set	Sampling	Quality Criterion	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$
CAS(ME) 2	-	-	1.80	2.24	0.99	1.72	2.74	0.98	2.00	8.16	0.93
Celeb-DF (v2)	-	-	11.94	13.92	0.36	37.34	40.82	-0.14	23.12	28.45	0.03
CAS(ME) 2 + Celeb-DF (v2)	-	-	6.91	12.26	0.57	31.07	36.37	-0.30	8.58	15.28	0.39
	Random	-	4.04	12.02	0.64	7.86	15.02	0.39	6.20	15.20	0.56
	Top/Bottom $K$	COVER	0.46	0.60	1.00	2.56	10.07	0.72	1.69	6.76	0.88

Table 2: Comparison of quality criteria and sampling strategies for training data selection, evaluated across various unsupervised rPPG methods on the PURE dataset. The best performing combination for each metric is highlighted in bold, while the second-best is underlined.

Training Set	Sampling	Quality Criterion	Gideon21			Contrast-Phys			SiNC
Training Set	Sampling	Quality Criterion	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$
CAS(ME) 2	-	-	2.58	4.45	0.96	2.72	3.74	0.93	10.90	16.31	0.30
Celeb-DF (v2)	-	-	13.60	19.07	0.13	23.24	28.55	-0.08	16.23	21.27	0.11
CAS(ME) 2 + Celeb-DF (v2)	-	-	5.97	10.83	0.72	9.38	15.89	0.52	13.61	19.23	0.19
	Random	-	3.28	4.91	0.96	7.69	15.20	0.41	10.93	17.93	0.42
	Top/Bottom $K$	COVER	2.83	4.36	0.96	4.45	9.39	0.79	4.46	11.17	0.66

Table 3: Comparison of quality criteria and sampling strategies for training data selection, evaluated across various unsupervised rPPG methods on the BUAA dataset.

Training Set	Test Set	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$
PURE	PURE	1.00	1.40	0.99
CAS(ME) 2 + Celeb-DF (v2) + PURE	PURE	0.15	0.18	1.00
PURE	BUAA	2.43	3.25	0.97
CAS(ME) 2 + Celeb-DF (v2) + PURE	BUAA	1.76	1.98	1.00
BUAA	BUAA	1.93	2.14	0.99
CAS(ME) 2 + Celeb-DF (v2) + BUAA	BUAA	1.12	1.59	0.99

Table 4: Impact of augmenting training data with “in-the-wild” videos. The Contrast-Phys method is used for model training.

Frequency Consistency Weight	Spectral Correlation Weight	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$
✗	✗	0.82	1.47	1.00
✓	✗	0.52	0.79	1.00
✗	✓	0.57	0.90	1.00
✓	✓	0.47	0.74	1.00

Table 5: Ablation study on the components of the signal-level noise perception branch.

Ranking-based Filtering	Duration-aware Sampling	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$
✗	✗	0.71	1.60	0.99
✓	✗	0.70	1.22	1.00
✗	✓	0.49	0.82	1.00
✓	✓	0.47	0.74	1.00

Table 6: Ablation study on the components of the TAS strategy.

$\alpha$	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$
0.0	6.91	14.06	0.49
0.2	5.59	12.11	0.63
0.4	4.39	10.81	0.70
0.6	1.52	6.54	0.89
0.8	0.47	0.74	1.00
1.0	0.78	1.15	1.00

Table 7: Ablation study on the fusion weight

\alpha

Figure 3: Prompt for Qwen3-VL to assess scene-level quality.

$M$	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$
1	0.91	1.30	0.99
3	0.78	1.17	0.99
5	0.57	1.12	1.00
7	0.47	0.74	1.00

Table 8: Ablation study on the fusion size

M

MLLM	MAE $\downarrow$	RMSE $\downarrow$	R $\uparrow$
Qwen3-VL	0.46 $\pm$ 0.02	0.77 $\pm$ 0.06	1.00 $\pm$ 0.00
Gemini 3.0 Pro	0.52 $\pm$ 0.05	0.73 $\pm$ 0.08	1.00 $\pm$ 0.00
Kimi K2.5	0.55 $\pm$ 0.09	0.86 $\pm$ 0.11	1.00 $\pm$ 0.00

Table 9: Generalization and stability analysis across different MLLMs.

Figure 4: Distribution of quality scores over the CAS(ME) 2 and Celeb-DF (v2) datasets.

Figure 5: Representative failure cases. (a) Environmental noise interference; (b) AI-synthesized synthetic faces.

Video Quality Assessment

Video quality assessment measures perceived visual and temporal fidelity in captured, compressed, streamed, and generated videos.

See all 52 papers

LQ-rPPG: A Label-Quantized Coarse-to-Fine Learning Framework for Remote Physiological Measurement
May 22, 2026Jun Seong Lee, Samyeul Noh, Changki Sung +1Remote Photoplethysmography Learning with Noisy Labels
Electronics and Telecommunications Research Institute, 218 Gajeong-ro, Yuseong-gu, Daejeon 34129, Republic of Korea · School of Electrical Engineering, Korea Advanced Institute of Science and Technology, 291 Daehak-ro, Yuseong-gu, Daejeon 34141, Republic of Korea
FCUS-rPPG: A Fast-Converging Unsupervised Framework for Remote Photoplethysmography via Gradient Oscillation Suppression
Jun 2, 2026Jiajie Li, Yu Liu, Rencheng Song +2Remote Photoplethysmography Unsupervised Representation Learning
Department of Biomedical Engineering, and also with Anhui Province Key Laboratory of Measuring Theory and Precision Instrument, Hefei University of Technology, Hefei 230009, China · Department of Electronic Engineering and Information Science, University of Science and Technology of China, Hefei 230026, China
StreamPPG: Low-Latency rPPG Estimation via Consistent Privileged Learning
Jun 22, 2026Yiming Li, Yihan Yang, Yuguang Chu +4Remote Photoplethysmography Photoplethysmography
ARXIV

May 22, 2026J/K move · Enter open · S save

LQ-rPPG: A Label-Quantized Coarse-to-Fine Learning Framework for Remote Physiological Measurement

Jun Seong Lee, Samyeul Noh, Changki Sung, Hyun Myung

Electronics and Telecommunications Research Institute, 218 Gajeong-ro, Yuseong-gu, Daejeon 34129, Republic of Korea · School of Electrical Engineering, Korea Advanced Institute of Science and Technology, 291 Daehak-ro, Yuseong-gu, Daejeon 34141, Republic of Korea

Remote Photoplethysmography Learning with Noisy Labels Photoplethysmography

Open paper AskPDF arXiv

Remote photoplethysmography (rPPG) enables non-contact measurement of physiological signals from facial videos, offering strong potential for remote healthcare and daily health monitoring. Driven by this potential, various deep learning-based rPPG methods have been proposed to improve rPPG estimation. However, previous deep learning-based rPPG methods have paid little attention to the quality of training labels and their impact on model learning. Contact-based PPG signals used as training labels often contain noise and variability caused by motion artifacts, inconsistent sensor contact, and morphological distortions. Such label inconsistency can lead models to overfit to the label noise and variability and consequently degrade generalization performance. To address this issue, we propose LQ-rPPG, a label-quantized coarse-to-fine learning framework for robust rPPG estimation. LQ-rPPG consists of a label quantization module and a coarse-to-fine rPPG estimation model. The label quantization module transforms continuous PPG signals into multi-bit quantized pseudo labels with reduced noise and variability. The coarse-to-fine estimation model progressively refines rPPG signals under hierarchical supervision guided by the multi-bit pseudo labels. This design alleviates overfitting to label-specific variations and enables the model to learn structured and consistent representations. As a result, LQ-rPPG achieves robust and generalizable rPPG estimation even under challenging conditions. Experiments on multiple benchmark datasets demonstrate that LQ-rPPG achieves strong performance in both intra- and cross-dataset evaluations, while reducing parameters and multiply-accumulate operations by 88% and 29%, respectively, and increasing throughput by 191%. The code is available at https://github.com/Anonymous-repo-code/LQ-rPPG.

Abstract

Video Quality Assessment

Explore similar work

LQ-rPPG: A Label-Quantized Coarse-to-Fine Learning Framework for Remote Physiological Measurement

FCUS-rPPG: A Fast-Converging Unsupervised Framework for Remote Photoplethysmography via Gradient Oscillation Suppression

StreamPPG: Low-Latency rPPG Estimation via Consistent Privileged Learning