Semi-rPPG: Semi-Supervised Remote Physiological Measurement with Curriculum Pseudo-Labeling

Bingjie Wu; Chaoqi Luo; Rick Siow Mong Goh; Wei Liu; Yiping Xie; Yong Liu; Zitong Yu

REVIEW 5 major objections 7 minor 50 references

Semi-rPPG: Semi-Supervised Remote Physiological Measurement with Curriculum Pseudo-Labeling

T0 review · 5 major / 7 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Semi-rPPG shows that 20% labeled face video can match fully supervised heart-rate accuracy when combined with unlabeled video.

desk verdict A plausible semi-supervised rPPG method, but the headline empirical claim is not yet established because of test-set tuning, modified baselines, and a broken equation. read the letter →

arxiv 2502.03855 v1 pith:7PHXSTL4 submitted 2025-02-06 cs.CV

Bingjie Wu , Zitong Yu , Yiping Xie , Wei Liu , Chaoqi Luo , Yong Liu , Rick Siow Mong Goh This is my paper · ORCID

classification cs.CV

keywords remotephotoplethysmographysemi-supervisedlearningcurriculumpseudo-labelingsignal-to-noiseratioconsistencyregularizationheartrateestimationfacialvideo

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Semi-rPPG is a training recipe for remote photoplethysmography (rPPG), the task of reading heart rate and blood-volume pulse from ordinary face videos. The paper's central claim is that when only 20% of training clips are labeled and the remaining 80% are unlabeled, the proposed combination of curriculum pseudo-labeling and a quasi-periodic consistency loss yields accuracy comparable to a fully supervised model: on the VIPL-HR benchmark, MAE 6.06 versus 5.82 BPM and RMSE 10.99 versus 10.71 BPM. The same recipe also improves cross-dataset transfer, reporting MAE 0.57 and RMSE 1.01 BPM when training on UBFC-rPPG with COHFACE as unlabeled data and testing on PURE. The reason to care is data economics: labeled physiological video is expensive and privacy-sensitive, while unlabeled face video is abundant, so a reliable semi-supervised route would let rPPG models scale beyond the small public datasets they currently train on.

What carries the argument

The load-bearing machinery is the SNR-ranked curriculum pseudo-labeling rule: a per-sample score computed from the model's own predicted power spectral density, $$\mathrm{SNR} = \frac{\sum_{i=\arg\max(P)-\$\Delta$ F}^{\arg\max(P)+\$\Delta$ F} P_i}{\sum_{i=a}^{b} P_i},$$ with $a=0.67$ Hz, $b=3$ Hz, and $\Delta F = 0.1$ Hz, combined with a selection step that keeps only the top $k = R N_{un}$ clips and a schedule $R = m + n e^{j/e_{total}}$ that raises $R$ from 0.2 to 0.8. This rule decides which unlabeled clips enter supervised training and therefore controls how much self-generated noise the model sees. The second machinery is the consistency loss $$\mathcal{L}_c = \mathrm{CE}\big(PSD(f_\$\theta$(x_s)), PSD(f_\$\theta$(x_w))\big),$$ between temporally reversed (strong) and temporally shifted (weak) clips, which supplies a label-free training signal that does not depend on pseudo-label correctness.

What would settle it

Take a held-out unlabeled split, rank clips by the SNR defined in the paper, and measure the heart-rate error of the pseudo-labels in the top and bottom quartiles; if the high-SNR quartile is not clearly more accurate than the low-SNR quartile, the curriculum's filtering premise collapses. A second direct check is to train the full pipeline with the SNR ranker replaced by a random selector using the same 20%-to-80% schedule; matching performance would show the ranker adds nothing.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that semi-supervised rPPG learning can recover fully supervised accuracy by combining two components: curriculum pseudo-labeling and quasi-periodic consistency regularization. Pseudo-labels are generated by the model itself for unlabeled clips; each clip's predicted blood-volume-pulse signal is scored by signal-to-noise ratio, defined as the power in a band around the PSD peak divided by the total band-limited power. Only the top fraction of clips, growing from 20% to 80% over training, is kept. In parallel, every clip is weakly augmented by a small temporal shift and strongly augmented by temporal reversal, and the consistency loss forces the power spectral densities of the two augmented predictions to match, because both share the same heart rate. The paper reports that this combination outperforms three classical semi-supervised baselines under the same 20/80 split and closes most of the gap to full supervision.

Load-bearing premise

The whole curriculum rests on the assumption that a high signal-to-noise ratio for the model's own predicted pulse spectrum really means a correct pseudo-label, and the paper does not measure pseudo-label error directly.

Editorial extensions

If this is right

The reported intra-dataset result implies that 80% of the labels in a supervised rPPG training set can be replaced by unlabeled clips without a meaningful accuracy penalty, reducing label collection from the dominant cost to a smaller fixed cost.
The cross-dataset result implies that unlabeled video from a third domain can serve as a bridge, reducing the domain gap between a labeled training set and a target test set.
The ablation results imply that both components matter: SNR-based curriculum selection beats an alternative power-ratio criterion by a large margin, and the time-reversal consistency loss beats two earlier self-supervised losses under the same curriculum.
The success of the increasing curriculum ratio (20% to 80%) implies that early training should be protected from most pseudo-labels, with the admission rate expanding only as the model stabilizes.
Near-parity with full supervision implies that the practical ceiling for rPPG accuracy may be set less by label quantity and more by the diversity of unlabeled video pools and the model backbone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the temporal-reverse consistency loss is not rPPG-specific; any quasi-periodic signal whose rate is read from the power spectrum admits the same label-preserving strong augmentation, so the recipe could transfer to respiration-rate or audio-based heart-rate regression.
Editorial inference: the SNR ranking assumes a single dominant spectral peak is the pulse signal; under motion or illumination artifacts a sharp but wrong peak might rank high, so an ensemble-disagreement or uncertainty-aware confidence filter would be a natural stronger alternative that the paper does not test.
Editorial inference: if unlabeled video from a different domain improves target-domain accuracy, then collecting large unlabeled face-video corpora from the target deployment environment may be a cheaper substitute for collecting labels in that environment, a data-collection strategy the paper does not pursue.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Let me give you the short version: the paper assembles known SSL ingredients in a sensible way for rPPG—SNR-ranked curriculum pseudo-labeling plus a temporal shift/reverse consistency loss—and it sets up a semi-supervised benchmark on four public datasets. That part is useful. But the central claim, that 20% labels match full supervision, is not backed by the evidence as reported.

The new bits are real. The SNR-based curriculum selection is a reasonable adaption of FlexMatch to quasi-periodic signals, and the consistency loss using temporal reversal (which preserves PSD) is a neat trick. The cross-dataset setup (UBFC labeled, COHFACE unlabeled, PURE test) is a good idea and the benchmark will be handy for the community. The ablations are extensive and the internal numbers are consistent with each other.

The problems are concentrated in the evaluation. First, the baseline SSL methods are not the original methods. FixMatch and FlexMatch are re-implemented with the proposed consistency loss, and self-training is modified with the SNR curriculum. So the comparison shows that the proposed components help on top of these methods, not that the method beats FixMatch or FlexMatch as published. Second, the hyperparameters (R schedule, lambda, pre-training epochs) appear to be chosen using the same Protocol 3 test setting that is later reported. There is no mention of a validation split for the ablations, and COHFACE has no validation split at all, so the reported gains likely include selection bias. Third, every number is a single run. The 0.28 BPM RMSE gap to fully supervised on VIPL-HR could easily be run-to-run noise. Fourth, Eq. (9) as written is wrong: with m=0.2, n=0.6, at j=0 you get R=0.8, not 20%, and at j=e_total you get R>1. The intended schedule probably works in the implementation, but this kind of error makes you worry about other details.

The SNR ranking assumption is also unverified—the paper assumes high SNR means accurate pseudo-label, but a sharp wrong peak under motion artifacts would break that. The ablation swapping SNR for IPR helps, but it is not an independent label-quality check.

So: for rPPG researchers, the benchmark and the consistency-loss idea are worth engaging with. The method is plausible and the paper deserves a serious referee. But the empirical claim should not be taken at face value until the authors rerun with validation-based hyperparameter selection, report multiple seeds, fix the equation, and compare against unmodified baselines. I would not cite it yet, but I'd send it to review and ask for those changes.

Referee Report

5 major / 7 minor

Summary. The paper proposes Semi-rPPG, a semi-supervised learning method for remote photoplethysmography (rPPG). The method combines curriculum pseudo-labeling, in which unlabeled facial video clips are ranked by the signal-to-noise ratio (SNR) of the model's predicted blood volume pulse (BVP) and only the top R fraction is selected for training, with a consistency regularization loss that encourages agreement between weakly augmented (time-shifted) and strongly augmented (time-reversed) clips. The method is evaluated on VIPL-HR, COHFACE, UBFC-rPPG, and PURE under intra-dataset and cross-dataset protocols. The central claim is that with 20% labeled data and 80% unlabeled data, Semi-rPPG achieves test performance comparable to fully supervised training on VIPL-HR (MAE 6.06 vs 5.82 BPM, RMSE 10.99 vs 10.71 BPM), and that in the UBFC-to-PURE cross-dataset setting with COHFACE as unlabeled data it achieves MAE 0.57 and RMSE 1.01 BPM. Extensive ablations compare selection ratios, SNR versus IPR criteria, different self-supervised losses, backbones, pre-training epochs, and the loss trade-off lambda.

Significance. If established, the result would be practically important because labeled rPPG data are expensive to collect, and a semi-supervised method that matches fully supervised performance with only 20% labels would reduce annotation burden. The paper also contributes the first semi-supervised rPPG benchmark spanning four public datasets and proposes a simple, physiologically motivated consistency loss based on time reversal and time shifting. The ablation study is broad and covers most design choices, and the reported tables are internally consistent. However, the empirical evidence does not yet reliably establish the headline claim: all results come from single runs with no error bars, the hyperparameter selection procedure is not separated from test-set reporting, the equation for the curriculum schedule cannot reproduce the stated 20%-80% range, and the baseline SSL methods are modified versions of the proposed method rather than the classical methods named in the text. No code is provided, which further limits reproducibility. These issues are fixable in revision.

major comments (5)

[III-B, Eq. (9)] Equation (9) cannot reproduce the claimed curriculum schedule. With m=0.2 and n=0.6, the formula R = m + n*exp(j/e_total) yields R=0.8 at j=0 and approximately 1.83 at j=e_total, not the stated 20% initial and 80% final values. Because the selection ratio R is the core mechanism of the curriculum pseudo-labeling and every reported experiment depends on it, the method as written is not reproducible. Please correct the schedule (e.g., a linear ramp R = 0.2 + 0.6*(j/e_total), or an exponential that starts at 0.2 and saturates at 0.8) and state the exact schedule used in all experiments.
[IV-B/IV-E, Tables VI-XI] The headline comparison is not protected from hyperparameter selection on the test set. The ablations in Section IV-E choose the ratio schedule, SNR criterion, consistency loss, pre-training epoch e_pre=1, and lambda=0.01 using Protocol 3; the text never states that these choices were made on the VIPL-HR validation split (which exists per Section IV-B) rather than on the same Protocol 3 test setting used for Table III. For COHFACE, Section IV-C explicitly states that no validation split is used, so the hyperparameters are presumably transferred from VIPL-HR, adding another layer of selection bias. Because the central claim is an empirical match to fully supervised performance (RMSE 10.99 vs 10.71), the manuscript must either demonstrate that all choices were made on validation data or report a nested validation procedure; otherwise the reported number is an optimistically selected estimate. The same concern applies to the cross-dataset results in Table V.
[Tables III-V] All reported metrics come from a single training run; no random seeds, standard deviations, or statistical tests are provided. The VIPL-HR claim of 'comparable to fully supervised' rests on a 0.28 BPM RMSE gap (10.99 vs 10.71), which is within plausible run-to-run variation for rPPG models. Please report results over at least three to five seeds (mean ± std) for the main comparisons and the key ablations, or provide a statistical significance test. Without this, the paper's central empirical conclusion is not reliably established.
[IV-C, Tables III-V] The comparison set is not what is claimed. Section IV-C says 'SSL-Self-training' is implemented with the proposed SNR curriculum pseudo-labeling (without consistency) and that FixMatch and FlexMatch are implemented with the proposed consistency loss Eq. (10). These are therefore ablated variants of Semi-rPPG, not the classical SSL methods cited in the text. The abstract and conclusion say Semi-rPPG outperforms 'three classical semi-supervised methods,' but the tables do not contain unmodified FixMatch, FlexMatch, or standard self-training. Please add the original baselines under the same protocol (e.g., FixMatch with a confidence threshold on the predicted class, FlexMatch with class-wise thresholds, and plain self-training), or rephrase the claim as a component-wise ablation.
[III-B, Eqs. (6)-(7)] The curriculum assumes that the SNR of the model's own predicted BVP is a trustworthy proxy for pseudo-label quality, but the paper provides no evidence for this correlation. A prediction with a sharp but incorrect spectral peak (e.g., caused by motion or illumination artifacts) could receive a high SNR and be selected, potentially amplifying the model's errors. Because the VIPL-HR validation split has ground-truth labels, this assumption can be tested directly: compute pseudo-labels on labeled validation clips, measure the correlation between SNR rank and BVP/HR error, and report failure cases. Please add such an analysis or an alternative quality measure; without it, the claimed benefit of curriculum selection over random selection is not mechanistically supported.

minor comments (7)

[Algorithm 1] In Algorithm 1, the epoch-1 loop samples 'xl, yl' from the merged dataset {Xnew, Ynew} and also samples 'xun' from {Xun}; since {Xnew, Ynew} already contains the selected pseudo-labeled clips, these clips can be used twice per epoch (once under the supervised loss and once under the consistency loss). Please clarify whether selected clips are removed from the unlabeled pool or whether this double use is intended.
[IV-C] The self-training baseline is cited to [47], which is a survey of self-training rather than a primary algorithmic source; please cite the original self-training algorithm or describe the exact implementation in enough detail to reproduce it.
[III-C] The consistency loss is said to be 'inspired by FlexTeacher method [42]', but reference [42] is FlexMatch; please correct the name or the reference.
[IV-B] The implementation section says the Adam optimizer is used 'with learning rates of 1e-4 and 1e-5' without specifying which rate applies to the supervised, pre-training, and semi-supervised stages or whether a schedule is used; please clarify.
[III-A, Eq. (3)] Equation (3) has missing parentheses in the denominator; the Pearson correlation formula should be written with explicit bracketing so that the numerator and denominator are unambiguous.
[IV-F1, Fig. 6] The sentence 'The minimum MAE for unlabelled data of Semi-rPPG is 6.06 BPM' repeats the test MAE from Table III; please clarify whether Fig. 6 reports pseudo-label error on unlabeled training clips or test-set MAE, and label the axis accordingly.
[Table IV] On COHFACE, SSL-Self Learning (MAE 6.66) is substantially worse than the partially supervised baseline (MAE 4.56); the text notes this but does not discuss why self-training fails on this dataset while helping on VIPL-HR. A brief analysis would be useful.

Circularity Check

0 steps flagged · score 2.0 of 10

No equation-level circularity: the semi-supervised loss terms compare the model to ground truth or to augmentations of the same input, and the headline claim is evaluated on external public benchmarks; minor self-citations appear but are not load-bearing.

full rationale

The claimed derivation chain is not circular. The supervised loss (Eqs. 1-4) compares the predicted BVP and heart-rate class against ground-truth labels; the consistency loss (Eq. 10) compares two augmented views of the same clip rather than regressing toward a fitted output; and the curriculum pseudo-labeling (Eqs. 5-9) selects unlabeled clips by ranking the SNR of the model's own predictions, which is a standard self-training property rather than a definitional tie between the input and the reported outcome. The central empirical claim (Table III: Semi-rPPG MAE 6.06, RMSE 10.99 vs. fully supervised MAE 5.82, RMSE 10.71) is evaluated on public datasets and against external SSL baselines, so it does not reduce by construction to fitted parameters or to a self-citation chain. Some self-citations are present, e.g., [33] for the classification-loss protocol and prior rPPG papers by co-authors, but the central result does not depend on an unverified self-citation: the baselines are external and the PhysNet backbone is a published architecture. The paper's ablations select hyperparameters such as lambda, R, and pre-training epochs, and no repeated-seed statistics are reported; additionally, Eq. (9) does not literally implement the stated 20%-80% ratio schedule. These are correctness and reproducibility risks, not circularity under the definitions used here. Therefore no circular step is identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central method rests on an empirical SNR proxy, an augmentation invariance assumption, and a cross-domain transfer assumption. No new physical entity is introduced. Several free parameters, especially the R schedule and lambda, are tuned against the reported evaluation protocol rather than derived from theory.

free parameters (4)

R-curriculum coefficients (m, n) = m = 0.2, n = 0.6
Eq. (9) sets the pseudo-label selection ratio R from 20% to 80%. These coefficients are not derived and the ablation in Table VI shows the R schedule strongly affects VIPL-HR results.
lambda (loss trade-off) = 0.01
Eq. (4) balances cross-entropy and Pearson losses. Table XI shows lambda = 0.01 performs best on the reported protocol, with no independent optimization procedure.
pre-training epochs e_pre = 1
Table X shows that more pre-training epochs hurt performance. The value of 1 is chosen by comparing test-protocol results rather than by a principled rule.
SNR peak window half-width Delta F = 0.1 Hz
Eq. (7) uses a 0.1 Hz window around the peak frequency to define SNR. This value is hand-set and no sensitivity analysis is provided.

assumptions (4)

domain assumption SNR of the model's predicted BVP pseudo-labels is a valid proxy for pseudo-label quality.
Section III-B, Eqs. (6)-(9): unlabeled data selection keeps or discards clips solely by ranking this SNR. No independent measure of label noise is used to validate the ranking.
domain assumption Temporal reversal of the video preserves the heart-rate information in the network output, so the PSD consistency loss is a well-posed training signal.
Section III-C: the loss assumes the network outputs should have matching PSDs for reversed and shifted clips because time reversal preserves the power spectrum of a quasi-periodic signal. It also assumes this cannot be satisfied by a degenerate constant output.
domain assumption The 0.67-3 Hz band limits fully cover the relevant heart-rate range for all datasets.
Section III-B and Eq. (11): heart-rate classes are derived from the PSD peak inside this band. Subjects with heart rates outside 40-180 BPM would be assigned to the wrong class.
domain assumption Unlabeled COHFACE videos carry transferable physiological signal structure for the UBFC-to-PURE task despite domain differences.
Section IV-D: the cross-dataset improvement is interpreted as evidence that SSL reduces domain gap. The result depends on COHFACE pseudo-labels being useful for the target PURE task.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Semi-rPPG: Semi-Supervised Remote Physiological Measurement with Curriculum Pseudo-Labeling." pith.science (2026). https://pith.science/paper/7PHXSTL4

@misc{pith2026250203855,
  author       = {Pith},
  title        = {Pith review of: Semi-rPPG: Semi-Supervised Remote Physiological Measurement with Curriculum Pseudo-Labeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7PHXSTL4}},
  note         = {Machine review of arXiv:2502.03855}
}

read the original abstract

Remote Photoplethysmography (rPPG) is a promising technique to monitor physiological signals such as heart rate from facial videos. However, the labeled facial videos in this research are challenging to collect. Current rPPG research is mainly based on several small public datasets collected in simple environments, which limits the generalization and scale of the AI models. Semi-supervised methods that leverage a small amount of labeled data and abundant unlabeled data can fill this gap for rPPG learning. In this study, a novel semi-supervised learning method named Semi-rPPG that combines curriculum pseudo-labeling and consistency regularization is proposed to extract intrinsic physiological features from unlabelled data without impairing the model from noises. Specifically, a curriculum pseudo-labeling strategy with signal-to-noise ratio (SNR) criteria is proposed to annotate the unlabelled data while adaptively filtering out the low-quality unlabelled data. Besides, a novel consistency regularization term for quasi-periodic signals is proposed through weak and strong augmented clips. To benefit the research on semi-supervised rPPG measurement, we establish a novel semi-supervised benchmark for rPPG learning through intra-dataset and cross-dataset evaluation on four public datasets. The proposed Semi-rPPG method achieves the best results compared with three classical semi-supervised methods under different protocols. Ablation studies are conducted to prove the effectiveness of the proposed methods.

Figures

Figures reproduced from arXiv: 2502.03855 by the authors.

**Figure 2.** The overall framework of the proposed Semi-rPPG. (a) In the supervised stage, the model is trained on labeled data using supervised loss and consistency regularization. (b) During the pseudo-labeling stage, the trained model assigns preliminary pseudo-labels to the unlabeled data. Next, the signal-to-noise ratio (SNR) is computed and sorted for all unlabeled samples. Top high-quality pseudo-labels are then adaptivel… view at source ↗

**Figure 3.** An illustration of calculating heart rate from blood volume pulse signal (BVP): First, the BVP signal is converted into power spectral density (PSD) through Fast Fourier Transform; Next, the PSD is categorized into classes ranging from 0 to 140, which correspond to heart rates of 40 to 180 beat per minute (BPM); Finally, the heart rate is determined by reverse mapping the class. Algorithm 1: Semi-rPPG Method Input: … view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Illustration of training datasets: the four datasets [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 6.** Figure 6: MAE of unlabelled data with the training epochs (a) Visualization of predicted blood volume pulse (BVP) (b) Visualization of predicted power density distribution (PSD) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 5.** Figure 5: The variations of SNR and IPR across training epochs [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

**Figure 7.** Figure 7: Visualization of the predicted and ground-truth BVP [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: Visualization of learned features. The brighter area [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 31 canonical work pages

[1]

Facial-video-based physiological signal measurement: Recent advances and affective applications,

Z. Yu, X. Li, and G. Zhao, “Facial-video-based physiological signal measurement: Recent advances and affective applications,” IEEE Signal Processing Magazine, vol. 38, no. 6, pp. 50–58, 2021

work page 2021
[2]

A robust real time system for remote heart rate measurement via camera,

D. N. Tran, H. Lee, and C. Kim, “A robust real time system for remote heart rate measurement via camera,” in 2015 IEEE International Conference on Multimedia and Expo (ICME) . IEEE, 2015, pp. 1–6

work page 2015
[3]

Robust heart rate estimation with spatial–temporal attention network from facial videos,

M. Hu, F. Qian, X. Wang, L. He, D. Guo, and F. Ren, “Robust heart rate estimation with spatial–temporal attention network from facial videos,” IEEE Transactions on Cognitive and Developmental Systems , vol. 14, no. 2, pp. 639–647, 2021

work page 2021
[4]

Deep super-resolution network for rppg information recovery and noncontact heart rate estimation,

Z. Yue, S. Ding, S. Yang, H. Yang, Z. Li, Y . Zhang, and Y . Li, “Deep super-resolution network for rppg information recovery and noncontact heart rate estimation,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–11, 2021

work page 2021
[5]

Time–frequency learning framework for rppg signal estimation using scalogram-based feature map of facial video data,

M. Das, M. K. Bhuyan, and L. Sharma, “Time–frequency learning framework for rppg signal estimation using scalogram-based feature map of facial video data,” IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–10, 2023

work page 2023
[6]

Eta-rppgnet: Effective time-domain attention network for remote heart rate measure- ment,

M. Hu, F. Qian, D. Guo, X. Wang, L. He, and F. Ren, “Eta-rppgnet: Effective time-domain attention network for remote heart rate measure- ment,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–12, 2021

2021
[7]

A facial-image- based blood pressure measurement system without calibration,

B.-F. Wu, B.-J. Wu, B.-R. Tsai, and C.-P. Hsu, “A facial-image- based blood pressure measurement system without calibration,” IEEE Transactions on Instrumentation and Measurement , vol. 71, pp. 1–13, 2022

2022
[8]

Assess- ment of deep learning based blood pressure prediction from ppg and rppg signals,

F. Schrumpf, P. Frenzel, C. Aust, G. Osterhoff, and M. Fuchs, “Assess- ment of deep learning based blood pressure prediction from ppg and rppg signals,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 3820–3830

work page 2021

Show all 50 references

[9]

Camera-based blood pressure estimation via windkessel model and waveform features,

B.-J. Wu, B.-F. Wu, and C.-P. Hsu, “Camera-based blood pressure estimation via windkessel model and waveform features,” IEEE Trans- actions on Instrumentation and Measurement , vol. 72, pp. 1–13, 2022

2022
[10]

Contactless spo 2 detection from face using consumer camera,

L. Zhu, K. Vatanparvar, M. Gwak, J. Kuang, and A. Gao, “Contactless spo 2 detection from face using consumer camera,” in 2022 IEEE-EMBS International Conference on Wearable and Implantable Body Sensor Networks (BSN). IEEE, 2022, pp. 1–4

2022
[11]

Weakly supervised rppg estimation for respiratory rate estimation,

J. Du, S.-Q. Liu, B. Zhang, and P. C. Yuen, “Weakly supervised rppg estimation for respiratory rate estimation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 2391–2397

2021
[12]

Video-based real-time monitoring for heart rate and respiration rate,

M. Alnaggar, A. I. Siam, M. Handosa, T. Medhat, and M. Rashad, “Video-based real-time monitoring for heart rate and respiration rate,” Expert Systems with Applications , vol. 225, p. 120135, 2023

2023
[13]

Atrial fibrillation detection from face videos by fusing subtle variations,

J. Shi, I. Alikhani, X. Li, Z. Yu, T. Sepp ¨anen, and G. Zhao, “Atrial fibrillation detection from face videos by fusing subtle variations,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 30, no. 8, pp. 2781–2795, 2019

2019
[14]

Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video,

X. Niu, H. Han, S. Shan, and X. Chen, “Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video,” in Computer Vision–ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V 14 . Spr...

2018
[15]

Non-contact video-based pulse rate measurement on a mobile service robot,

R. Stricker, S. M ¨uller, and H.-M. Gross, “Non-contact video-based pulse rate measurement on a mobile service robot,” in The 23rd IEEE Inter- national Symposium on Robot and Human Interactive Communication . IEEE, 2014, pp. 1056–1062

2014
[16]

Un- supervised skin tissue segmentation for remote photoplethysmography,

S. Bobbia, R. Macwan, Y . Benezeth, A. Mansouri, and J. Dubois, “Un- supervised skin tissue segmentation for remote photoplethysmography,” Pattern Recognition Letters, vol. 124, pp. 82–90, 2019

2019
[17]

A reproducible study on remote heart rate measurement,

G. Heusch, A. Anjos, and S. Marcel, “A reproducible study on remote heart rate measurement,” arXiv preprint arXiv:1709.00962 , 2017

2017 arXiv
[18]

Celebv-hq: A large-scale video facial attributes dataset,

H. Zhu, W. Wu, W. Zhu, L. Jiang, S. Tang, L. Zhang, Z. Liu, and C. C. Loy, “Celebv-hq: A large-scale video facial attributes dataset,” in European conference on computer vision. Springer, 2022, pp. 650–667

2022
[19]

Robust pulse rate from chrominance-based rppg,

G. De Haan and V . Jeanne, “Robust pulse rate from chrominance-based rppg,” IEEE transactions on biomedical engineering , vol. 60, no. 10, pp. 2878–2886, 2013

2013
[20]

Algorithmic principles of remote ppg,

W. Wang, A. C. Den Brinker, S. Stuijk, and G. De Haan, “Algorithmic principles of remote ppg,” IEEE Transactions on Biomedical Engineer- ing, vol. 64, no. 7, pp. 1479–1491, 2016. IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT 11

2016
[21]

Realistic pulse waveforms estimation via contrastive learning in remote photoplethysmography,

B. Dong, Y . Liu, K. Yang, and J. Cao, “Realistic pulse waveforms estimation via contrastive learning in remote photoplethysmography,” IEEE Transactions on Instrumentation and Measurement , 2024

2024
[22]

Dual-bridging with adversarial noise generation for domain adaptive rppg estimation,

J. Du, S.-Q. Liu, B. Zhang, and P. C. Yuen, “Dual-bridging with adversarial noise generation for domain adaptive rppg estimation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 10 355–10 364

2023
[23]

Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement,

Z. Yu, W. Peng, X. Li, X. Hong, and G. Zhao, “Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement,” in ICCV, 2019, pp. 151–160

2019
[24]

Physformer++: Facial video-based physiological measurement with slowfast temporal difference transformer,

Z. Yu, Y . Shen, J. Shi, H. Zhao, Y . Cui, J. Zhang, P. Torr, and G. Zhao, “Physformer++: Facial video-based physiological measurement with slowfast temporal difference transformer,” International Journal of Computer Vision, vol. 131, no. 6, pp. 1307–1330, 2023

2023
[25]

Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss,

H. Wang, E. Ahn, and J. Kim, “Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 2, 2022, pp. 2431–2439

2022
[26]

Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast,

Z. Sun and X. Li, “Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast,” in European Conference on Computer Vision . Springer, 2022, pp. 492–510

2022
[27]

The way to my heart is through contrastive learning: Remote photoplethysmography from unlabelled video,

J. Gideon and S. Stent, “The way to my heart is through contrastive learning: Remote photoplethysmography from unlabelled video,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 3995–4004

2021
[28]

Non-contrastive unsuper- vised learning of physiological signals from video,

J. Speth, N. Vance, P. Flynn, and A. Czajka, “Non-contrastive unsuper- vised learning of physiological signals from video,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14 464–14 474

2023
[29]

rppg-mae: Self- supervised pretraining with masked autoencoders for remote physiolog- ical measurements,

X. Liu, Y . Zhang, Z. Yu, H. Lu, H. Yue, and J. Yang, “rppg-mae: Self- supervised pretraining with masked autoencoders for remote physiolog- ical measurements,” IEEE Transactions on Multimedia , 2024

2024
[30]

Self-supervised rgb-nir fusion video vision transformer framework for rppg estimation,

S. Park, B.-K. Kim, and S.-Y . Dong, “Self-supervised rgb-nir fusion video vision transformer framework for rppg estimation,” IEEE Trans- actions on Instrumentation and Measurement , vol. 71, pp. 1–10, 2022

2022
[31]

Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation,

X. Niu, S. Shan, H. Han, and X. Chen, “Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation,” IEEE Transactions on Image Processing , vol. 29, pp. 2409–2423, 2019

2019
[32]

Video-based remote physiological measurement via cross-verified feature disentangling,

X. Niu, Z. Yu, H. Han, X. Li, S. Shan, and G. Zhao, “Video-based remote physiological measurement via cross-verified feature disentangling,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16 . Springer, 2020, pp. 295–310

2020
[33]

Physformer: Facial video-based physiological measurement with temporal difference transformer,

Z. Yu, Y . Shen, J. Shi, H. Zhao, P. H. Torr, and G. Zhao, “Physformer: Facial video-based physiological measurement with temporal difference transformer,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 4186–4196

2022
[34]

Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation,

A. Das, H. Lu, H. Han, A. Dantcheva, S. Shan, and X. Chen, “Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation,” in 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021) . IEEE, 2021, pp. 01–08

2021
[35]

Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks,

Z. Yu, X. Li, and G. Zhao, “Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks,” in BMVC, 2019, p. 277

2019
[36]

Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,

A. Tarvainen and H. Valpola, “Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,” Advances in neural information processing systems , vol. 30, 2017

2017
[37]

Rethinking pseudo labels for semi-supervised object detection,

H. Li, Z. Wu, A. Shrivastava, and L. S. Davis, “Rethinking pseudo labels for semi-supervised object detection,” in Proceedings of the AAAI conference on artificial intelligence, vol. 36, no. 2, 2022, pp. 1314–1322

2022
[38]

Semi- supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank,

I. Alonso, A. Sabater, D. Ferstl, L. Montesano, and A. C. Murillo, “Semi- supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 8219–8228

2021
[39]

Videossl: Semi- supervised learning for video classification,

L. Jing, T. Parag, Z. Wu, Y . Tian, and H. Wang, “Videossl: Semi- supervised learning for video classification,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision , 2021, pp. 1110–1119

2021
[40]

Pseudo- label correction and learning for semi-supervised object detection,

Y . He, W. Chen, K. Liang, Y . Tan, Z. Liang, and Y . Guo, “Pseudo- label correction and learning for semi-supervised object detection,”arXiv preprint arXiv:2303.02998, 2023

2023 arXiv
[41]

Fixmatch: Simplifying semi- supervised learning with consistency and confidence,

K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “Fixmatch: Simplifying semi- supervised learning with consistency and confidence,” Advances in neural information processing systems , vol. 33, pp. 596–608, 2020

2020
[42]

Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling,

B. Zhang, Y . Wang, W. Hou, H. Wu, J. Wang, M. Okumura, and T. Shi- nozaki, “Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling,” Advances in Neural Information Processing Systems , vol. 34, pp. 18 408–18 419, 2021

2021
[43]

Srppg: Semi-supervised adversarial learning for remote photoplethysmography with noisy data,

Z. Hasan, A. Z. M. Faridee, M. Ahmed, S. Ayyanar, and N. Roy, “Srppg: Semi-supervised adversarial learning for remote photoplethysmography with noisy data,” in 2023 IEEE International Conference on Smart Computing (SMARTCOMP). IEEE, 2023, pp. 25–32

2023
[44]

Contactless pulse estimation leveraging pseudo labels and self-supervision,

Z. Li and L. Yin, “Contactless pulse estimation leveraging pseudo labels and self-supervision,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 20 588–20 597

2023
[45]

Joint face detection and facial expression recog- nition with mtcnn,

J. Xiang and G. Zhu, “Joint face detection and facial expression recog- nition with mtcnn,” in 2017 4th international conference on information science and control engineering (ICISCE) . IEEE, 2017, pp. 424–427

2017
[46]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017

2017 arXiv
[47]

Self-training: A survey,

M.-R. Amini, V . Feofanov, L. Pauletto, E. Devijver, and Y . Maximov, “Self-training: A survey,” arXiv preprint arXiv:2202.12040 , 2022

2022 arXiv
[48]

rppg-toolbox: Deep remote ppg toolbox,

X. Liu, G. Narayanswamy, A. Paruchuri, X. Zhang, J. Tang, Y . Zhang, R. Sengupta, S. Patel, Y . Wang, and D. McDuff, “rppg-toolbox: Deep remote ppg toolbox,” Advances in Neural Information Processing Sys- tems, vol. 36, 2024

2024
[49]

Multi-task temporal shift attention networks for on-device contactless vitals measurement,

X. Liu, J. Fromm, S. Patel, and D. McDuff, “Multi-task temporal shift attention networks for on-device contactless vitals measurement,” Ad- vances in Neural Information Processing Systems , vol. 33, pp. 19 400– 19 411, 2020

2020
[50]

Deepphys: Video-based physiological mea- surement using convolutional attention networks,

W. Chen and D. McDuff, “Deepphys: Video-based physiological mea- surement using convolutional attention networks,” in Proceedings of the european conference on computer vision (ECCV) , 2018, pp. 349–365

2018

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Facial-video-based physiological signal measurement: Recent advances and affective applications,

Z. Yu, X. Li, and G. Zhao, “Facial-video-based physiological signal measurement: Recent advances and affective applications,” IEEE Signal Processing Magazine, vol. 38, no. 6, pp. 50–58, 2021

work page 2021

[2] [2]

A robust real time system for remote heart rate measurement via camera,

D. N. Tran, H. Lee, and C. Kim, “A robust real time system for remote heart rate measurement via camera,” in 2015 IEEE International Conference on Multimedia and Expo (ICME) . IEEE, 2015, pp. 1–6

work page 2015

[3] [3]

Robust heart rate estimation with spatial–temporal attention network from facial videos,

M. Hu, F. Qian, X. Wang, L. He, D. Guo, and F. Ren, “Robust heart rate estimation with spatial–temporal attention network from facial videos,” IEEE Transactions on Cognitive and Developmental Systems , vol. 14, no. 2, pp. 639–647, 2021

work page 2021

[4] [4]

Deep super-resolution network for rppg information recovery and noncontact heart rate estimation,

Z. Yue, S. Ding, S. Yang, H. Yang, Z. Li, Y . Zhang, and Y . Li, “Deep super-resolution network for rppg information recovery and noncontact heart rate estimation,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–11, 2021

work page 2021

[5] [5]

Time–frequency learning framework for rppg signal estimation using scalogram-based feature map of facial video data,

M. Das, M. K. Bhuyan, and L. Sharma, “Time–frequency learning framework for rppg signal estimation using scalogram-based feature map of facial video data,” IEEE Transactions on Instrumentation and Measurement, vol. 72, pp. 1–10, 2023

work page 2023

[6] [6]

Eta-rppgnet: Effective time-domain attention network for remote heart rate measure- ment,

M. Hu, F. Qian, D. Guo, X. Wang, L. He, and F. Ren, “Eta-rppgnet: Effective time-domain attention network for remote heart rate measure- ment,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–12, 2021

2021

[7] [7]

A facial-image- based blood pressure measurement system without calibration,

B.-F. Wu, B.-J. Wu, B.-R. Tsai, and C.-P. Hsu, “A facial-image- based blood pressure measurement system without calibration,” IEEE Transactions on Instrumentation and Measurement , vol. 71, pp. 1–13, 2022

2022

[8] [8]

Assess- ment of deep learning based blood pressure prediction from ppg and rppg signals,

F. Schrumpf, P. Frenzel, C. Aust, G. Osterhoff, and M. Fuchs, “Assess- ment of deep learning based blood pressure prediction from ppg and rppg signals,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 3820–3830

work page 2021