pith. sign in

arxiv: 2606.03802 · v1 · pith:MBTG6PD2new · submitted 2026-06-02 · 💻 cs.CV

Template Collapse and Information-Theoretic Limits in Camera rPPG Pulse Morphology Restoration

Achraf Ben Ahmed This is my paper

Pith reviewed 2026-06-28 10:38 UTC · model grok-4.3

classification 💻 cs.CV
keywords rPPGremote photoplethysmographypulse morphologytemplate collapsearterial stiffnesscamera-based monitoringinformation limitswaveform reconstruction
0
0 comments X

The pith

No architecture recovers subject-specific pulse morphology from single-cycle camera rPPG due to template collapse.

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

The paper tests whether single-cycle rPPG waveforms captured by consumer face cameras contain recoverable subject-specific arterial morphology that could indicate individual arterial stiffness. It evaluates 16 architectures from six families across 153 subjects in three datasets and introduces cross-subject Pearson r to measure whether outputs reflect individual traits or collapse to shared templates. Every model produces outputs with cross-subject correlations between 0.773 and 0.9999, well above the ground-truth ceiling of 0.601. Supervised contrastive learning reaches log N of 4.844 and supplies the strongest evidence that the input signals lack extractable discriminative structure. The result indicates that consumer cameras cannot deliver individual-level morphology information from rPPG.

Core claim

No architecture recovered subject-specific morphology (cross-subject r range 0.773--0.9999; ground-truth ceiling 0.601). Supervised Contrastive (SupCon) converged to log N = 4.844, constituting the strongest available empirical evidence that no discriminative morphological structure is extractable from single-cycle rPPG by the encoder families tested. The VAE decoder restores population-level harmonic content absent from the rPPG input (H2/H1: 0.310 output vs. 0.275 input), generalising zero-shot to UBFC (r = +0.708); a directional hallucination gap (p = 0.150) suggests partial signal reading.

What carries the argument

Cross-subject Pearson r diagnostic that separates template collapse from subject-specific recovery in rPPG waveform reconstruction tasks.

If this is right

  • Anti-collapse objectives cannot succeed when the input signal itself carries no discriminative morphological structure.
  • VAE-style decoders can restore population-level harmonic content and generalise zero-shot across datasets even without subject-specific recovery.
  • Cross-subject r becomes a required diagnostic for any waveform reconstruction benchmark claiming morphology restoration.
  • Consumer-camera rPPG is limited to population-level cardiovascular signals rather than individual arterial morphology.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Multi-cycle or multi-modal inputs may be required to access subject-specific morphology that single-cycle rPPG lacks.
  • Remote cardiovascular monitoring applications that assume camera rPPG can supply arterial stiffness biomarkers will need re-evaluation.
  • New encoder families can be screened with the cross-subject r metric to test whether they overcome the observed information limit.

Load-bearing premise

The 16 architectures spanning six families evaluated on the three datasets are representative enough to support the conclusion that no discriminative morphological structure is extractable from single-cycle rPPG.

What would settle it

An architecture that produces cross-subject Pearson r near the ground-truth ceiling of 0.601 while preserving high fidelity to individual subject waveforms on the same datasets would falsify the claim.

Figures

Figures reproduced from arXiv: 2606.03802 by Achraf Ben Ahmed.

Figure 1
Figure 1. Figure 1: Ground-truth morphological diversity versus VAE-Base predictions [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SupCon training curves for all six architectural variants (colours). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean H2/H1 harmonic ratio per architecture at three signal stages: [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative illustration of population-level waveform restoration. Best [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Template collapse: inter-subject standard deviation of predicted H2/H1 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Per-subject r vs. cross-subject r for all 16 architectures. Higher per-subject r (horizontal axis) is better; lower cross-subject r (vertical axis) indicates less collapse. The green dashed line marks the GT ceiling (rcross = 0.601); no architecture reaches it. RGB-Window achieves the highest per-subject r = 0.903 yet sits at near-total collapse (rcross = 0.996), demonstrating that per-subject r alone cann… view at source ↗
read the original abstract

Objective: Consumer face camera remote photoplethysmography (rPPG) enables passive cardiovascular monitoring, but whether single-cycle waveform morphology encoding arterial stiffness biomarkers is recoverable from this measurement has not been characterised. Methods: We evaluated 16 architectures spanning six families on 153 subjects across three datasets, introducing cross-subject Pearson r to distinguish subject-specific recovery from template collapse. Results: No architecture recovered subject-specific morphology (cross-subject r range 0.773--0.9999; ground-truth ceiling 0.601). Supervised Contrastive (SupCon) converged to log N = 4.844, constituting the strongest available empirical evidence that no discriminative morphological structure is extractable from single-cycle rPPG by the encoder families tested. The VAE decoder restores population-level harmonic content absent from the rPPG input (H2/H1: 0.310 output vs. 0.275 input), generalising zero-shot to UBFC (r = +0.708); a directional hallucination gap (p = 0.150) suggests partial signal reading. Anti-collapse objectives fail when input carries no discriminative structure. Significance: Consumer cameras cannot encode individual arterial morphology; cross-subject r is a necessary collapse diagnostic for waveform reconstruction benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript evaluates 16 architectures from six families on single-cycle rPPG signals from 153 subjects across three datasets. It introduces cross-subject Pearson r as a diagnostic for template collapse versus subject-specific morphology recovery and reports that all tested models fail to recover subject-specific structure (cross-subject r range 0.773--0.9999 versus ground-truth ceiling 0.601), with Supervised Contrastive learning reaching log N = 4.844. A VAE decoder is shown to restore population-level harmonics and generalize zero-shot; the paper concludes that consumer cameras cannot encode individual arterial morphology.

Significance. If the empirical negative result holds under the tested conditions, the work supplies concrete evidence of information-theoretic limits on morphology restoration from camera rPPG, with the cross-subject r metric and the observation of VAE hallucination of absent harmonics constituting useful contributions. The breadth of the 16-architecture sweep strengthens the case that collapse is not an artifact of a single model family.

major comments (3)
  1. [Abstract (Significance)] Abstract, Significance: The unqualified statement that 'Consumer cameras cannot encode individual arterial morphology' extrapolates from negative results on a finite set of 16 architectures without any coverage analysis, theoretical motivation, or argument that the six families are representative of encoders that could extract morphology if present; this generalization is load-bearing for the central claim.
  2. [Results] Results: The reported cross-subject r range (0.773--0.9999) and SupCon log N = 4.844 are presented as the strongest available evidence, yet the manuscript provides no visible full results tables, exact preprocessing pipeline, model implementation details, or statistical testing procedures, preventing independent verification that these metrics demonstrate absence of extractable structure rather than implementation artifacts.
  3. [Results] Methods/Results: The ground-truth ceiling of 0.601 is used to benchmark the observed r values, but no derivation or section explains how this ceiling was computed from the data or why it constitutes an upper bound on recoverable morphology; this quantity is central to interpreting the collapse claim.
minor comments (2)
  1. [Abstract] The term 'log N convergence' and the precise definition of N are used in the abstract and results without an explicit equation or paragraph defining the metric in the main text.
  2. [Results] The p = 0.150 value for the directional hallucination gap is reported but lacks a description of the underlying statistical test or how the gap was quantified.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for these constructive comments. We respond point-by-point below, indicating revisions where the manuscript can be strengthened for clarity and completeness.

read point-by-point responses

  1. Referee: [Abstract (Significance)] Abstract, Significance: The unqualified statement that 'Consumer cameras cannot encode individual arterial morphology' extrapolates from negative results on a finite set of 16 architectures without any coverage analysis, theoretical motivation, or argument that the six families are representative of encoders that could extract morphology if present; this generalization is load-bearing for the central claim.

    Authors: We agree the Significance statement requires qualification to match the empirical scope. The six families were chosen to cover dominant paradigms in the rPPG and time-series literature (CNN, RNN, Transformer/attention, contrastive, variational, and hybrid). In revision we will change the Significance sentence to end with 'by the encoder families tested' and add a brief paragraph in Discussion noting the selection rationale while acknowledging that novel architectures could in principle behave differently. revision: yes

  2. Referee: [Results] Results: The reported cross-subject r range (0.773--0.9999) and SupCon log N = 4.844 are presented as the strongest available evidence, yet the manuscript provides no visible full results tables, exact preprocessing pipeline, model implementation details, or statistical testing procedures, preventing independent verification that these metrics demonstrate absence of extractable structure rather than implementation artifacts.

    Authors: We accept that full transparency is required. The current text reports summary statistics; the revision will move the complete per-architecture table, the exact preprocessing steps (filtering, cycle segmentation, normalization), model hyper-parameters, and the statistical procedures used to obtain the reported ranges and log N into the supplementary material. revision: yes

  3. Referee: [Results] Methods/Results: The ground-truth ceiling of 0.601 is used to benchmark the observed r values, but no derivation or section explains how this ceiling was computed from the data or why it constitutes an upper bound on recoverable morphology; this quantity is central to interpreting the collapse claim.

    Authors: The value 0.601 is the mean pairwise cross-subject Pearson correlation computed on the ground-truth single-cycle PPG waveforms pooled across the three datasets; it supplies an empirical upper bound on inter-subject morphological similarity. We will insert a short subsection in Methods titled 'Ground-Truth Inter-Subject Correlation Ceiling' that derives the quantity and explains its role as benchmark. revision: yes

Circularity Check

0 steps flagged

No circularity: results are direct empirical computations on held-out data

full rationale

The paper reports cross-subject Pearson r, log N convergence, and harmonic ratios computed directly from running 16 architectures on three external datasets with held-out subjects. No equations define a quantity by fitting and then reuse it as a 'prediction'; no self-citation chain justifies the central negative claim; the evaluation stands on external benchmarks rather than internal definitions. This is self-contained empirical work with no load-bearing reductions of the enumerated kinds.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper is an empirical machine-learning study that relies on standard statistical definitions and domain assumptions about model coverage rather than introducing new free parameters, axioms, or entities.

axioms (2)
  • standard math Pearson correlation is an appropriate measure for quantifying waveform similarity and template collapse
    Used to define the cross-subject r diagnostic in the methods
  • domain assumption The 16 architectures from six families are sufficient to test whether any encoder can extract discriminative morphology
    Invoked to support the claim that no architecture recovered subject-specific morphology

pith-pipeline@v0.9.1-grok · 5752 in / 1410 out tokens · 37015 ms · 2026-06-28T10:38:13.637799+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 15 canonical work pages

  1. [1]

    Arterial stiffness and hypertension,

    H. Kim, “Arterial stiffness and hypertension,”Clinical Hypertension, vol. 29, no. 31, pp. 1–9, 2023, doi: 10.1186/s40885-023-00258-1

    work page doi:10.1186/s40885-023-00258-1 2023

  2. [2]

    C., Stuijk, S., & De Haan, G

    W. Wang, A. C. den Brinker, S. Stuijk, and G. de Haan, “Algorithmic principles of remote-PPG,”IEEE Trans. Biomed. Eng., vol. 64, no. 7, pp. 1479–1491, Jul. 2017, doi: 10.1109/TBME.2016.2609282

    work page doi:10.1109/tbme.2016.2609282 2017

  3. [3]

    Joint generative and contrastive learning for unsupervised person re- identification,

    H. Chen, Y . Wang, B. Lagadec, A. Dantcheva, and F. Bremond, “Joint generative and contrastive learning for unsupervised person re- identification,” inProc. IEEE/CVF CVPR, 2021, arXiv:2012.09071

    arXiv 2021

  4. [4]

    Motion-focused contrastive learning of video representations,

    R. Li, Y . Zhang, Z. Qiu, T. Yao, D. Liu, and T. Mei, “Motion-focused contrastive learning of video representations,” inProc. IEEE/CVF ICCV, 2021, arXiv:2201.04029

    arXiv 2021

  5. [5]

    sCL-ST: Su- pervised contrastive learning with semantic transformations for multiple lead ECG arrhythmia classification,

    D. Le, S. Truong, B. Patel, D. A. Adjeroh, and N. Le, “sCL-ST: Su- pervised contrastive learning with semantic transformations for multiple lead ECG arrhythmia classification,”IEEE J. Biomed. Health Inform., vol. 27, no. 6, pp. 2818–2828, 2023, doi: 10.1109/JBHI.2023.3246241

    work page doi:10.1109/jbhi.2023.3246241 2023

  6. [6]

    Functional emotion transformer for EEG-assisted cross-modal emotion recognition,

    W.-B. Jiang, Z. Li, W.-L. Zheng, and B.-L. Lu, “Functional emotion transformer for EEG-assisted cross-modal emotion recognition,” inProc. IEEE ICASSP, 2024

    2024

  7. [7]

    Self-supervised contrastive learning performs non-linear system identification,

    R. Gonz ´alez Laiz, T. Schmidt, and S. Schneider, “Self-supervised contrastive learning performs non-linear system identification,” inProc. ICLR, 2025, arXiv:2410.14673

    arXiv 2025

  8. [8]

    Sub- ject invariant contrastive learning for human activity recognition,

    Y . Yarici, K. Kokilepersaud, M. Prabhushankar, and G. AlRegib, “Sub- ject invariant contrastive learning for human activity recognition,” in IEEE 35th Int. Workshop Mach. Learn. Signal Process. (MLSP), 2025

    2025

  9. [9]

    Robust pulse rate from chrominance-based rPPG,

    G. de Haan and V . Jeanne, “Robust pulse rate from chrominance-based rPPG,”IEEE Trans. Biomed. Eng., vol. 60, no. 10, pp. 2878–2886, Oct. 2013, doi: 10.1109/TBME.2013.2266196

    work page doi:10.1109/tbme.2013.2266196 2013

  10. [10]

    Cross modal distillation for super- vision transfer,

    S. Gupta, J. Hoffman, and J. Malik, “Cross modal distillation for super- vision transfer,” inProc. IEEE/CVF CVPR, 2016, arXiv:1507.00448

    Pith/arXiv arXiv 2016

  11. [11]

    Non -contact, automated cardiac pulse measurements using video imaging and blind source separation

    M.-Z. Poh, D. J. McDuff, and R. W. Picard, “Non-contact, automated cardiac pulse measurements using video imaging and blind source separation,”Opt. Express, vol. 18, no. 10, pp. 10762–10774, 2010, doi: 10.1364/OE.18.010762

    work page doi:10.1364/oe.18.010762 2010

  12. [12]

    Understanding dimensional collapse in contrastive self-supervised learning,

    L. Jing, P. Vincent, Y . LeCun, and Y . Tian, “Understanding dimensional collapse in contrastive self-supervised learning,” inProc. ICLR, 2022, arXiv:2110.09348

    arXiv 2022

  13. [13]

    SyncGAN: Synchronize the latent space of cross-modal generative adversarial networks,

    W.-C. Chen, C.-W. Chen, and M.-C. Hu, “SyncGAN: Synchronize the latent space of cross-modal generative adversarial networks,” inProc. IEEE ICME, 2018, arXiv:1804.00410

    Pith/arXiv arXiv 2018

  14. [14]

    Cross-modal variational auto-encoder with distributed latent spaces and associators,

    D. U. Jo, B. Lee, J. Choi, H. Yoo, and J. Y . Choi, “Cross-modal variational auto-encoder with distributed latent spaces and associators,” arXiv:1905.12867, 2019

    Pith/arXiv arXiv 1905

  15. [15]

    Deep latent space learning for cross-modal mapping of audio and visual signals,

    S. Nawaz, M. K. Janjua, I. Gallo, A. Mahmood, and A. Calefati, “Deep latent space learning for cross-modal mapping of audio and visual signals,” inProc. IEEE DICTA, 2019, arXiv:1909.08685

    arXiv 2019

  16. [16]

    Preserving shape details of pulse signals for video-based blood pressure estimation,

    X. Han, X. Yang, S. Fang, Y . Chen, Q. Chen, L. Li, and R. Song, “Preserving shape details of pulse signals for video-based blood pressure estimation,”Biomed. Opt. Express, vol. 15, no. 4, pp. 2433–2450, 2024, doi: 10.1364/BOE.516388

    work page doi:10.1364/boe.516388 2024

  17. [17]

    Supervised contrastive learning,

    P. Khoslaet al., “Supervised contrastive learning,” inAdv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020, arXiv:2004.11362

    arXiv 2020

  18. [18]

    Joint contrastive learning with infinite possibilities,

    Q. Cai, Y . Wang, Y . Pan, T. Yao, and T. Mei, “Joint contrastive learning with infinite possibilities,” inAdv. Neural Inf. Process. Syst. (NeurIPS), vol. 33, 2020, arXiv:2009.14776

    arXiv 2020

  19. [19]

    Guided variational autoen- coder for speech enhancement with a supervised classifier,

    G. Carbajal, J. Richter, and T. Gerkmann, “Guided variational autoen- coder for speech enhancement with a supervised classifier,” inProc. IEEE ICASSP, 2021, pp. 681–685, arXiv:2102.06454

    arXiv 2021

  20. [20]

    Towards fine-grained visual rep- resentations by combining contrastive learning with image reconstruc- tion and attention-weighted pooling,

    J. Dippel, S. V ogler, and J. H ¨ohne, “Towards fine-grained visual rep- resentations by combining contrastive learning with image reconstruc- tion and attention-weighted pooling,” inICML 2021 Workshop: Self- Supervised Learning for Reasoning and Perception, 2021

    2021

  21. [21]

    Variational supervised contrastive learning,

    Z. Wang, J. Fan, T. Nguyen, H. Ji, and G. Liu, “Variational supervised contrastive learning,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2025, arXiv:2506.07413

    arXiv 2025

  22. [22]

    Photoplethysmography for the assessment of arterial stiffness,

    P. Karimpour, J. M. May, and P. A. Kyriacou, “Photoplethysmography for the assessment of arterial stiffness,”Sensors, vol. 23, no. 24, p. 9882, 2023, doi: 10.3390/s23249882

    work page doi:10.3390/s23249882 2023

  23. [23]

    Reduction of motion artifacts from photoplethysmography signals using learned convolutional sparse coding,

    G. Basso, X. Long, R. Haakma, and R. Vullings, “Reduction of motion artifacts from photoplethysmography signals using learned convolutional sparse coding,” arXiv:2508.10805, 2025

    arXiv 2025

  24. [24]

    SkinMap: Weighted full-body skin segmentation for robust remote photoplethysmography,

    Z. Malekiet al., “SkinMap: Weighted full-body skin segmentation for robust remote photoplethysmography,” arXiv:2510.05296, 2025

    arXiv 2025

  25. [25]

    Weighted combination and singular spectrum analysis based remote photoplethysmography pulse extraction in low-light en- vironments,

    L. Xiet al., “Weighted combination and singular spectrum analysis based remote photoplethysmography pulse extraction in low-light en- vironments,” arXiv:2503.03780, 2025

    arXiv 2025

  26. [26]

    FreqPhys: Repurposing implicit physiological frequency prior for robust remote photoplethysmography,

    W. Qianet al., “FreqPhys: Repurposing implicit physiological frequency prior for robust remote photoplethysmography,” arXiv:2604.00534, 2026

    arXiv 2026

  27. [27]

    ToTMNet: FFT-accelerated Toeplitz temporal mixing network for lightweight remote photoplethys- mography,

    V . Frants, S. Agaian, and K. Panetta, “ToTMNet: FFT-accelerated Toeplitz temporal mixing network for lightweight remote photoplethys- mography,” arXiv:2601.04159, 2026

    arXiv 2026

  28. [28]

    CP-PPG: Restoring photoplethysmography waveform morphology from poor skin contact,

    T. H. Phamet al., “CP-PPG: Restoring photoplethysmography waveform morphology from poor skin contact,”Sci. Rep., 2025, doi: 10.1038/s41598-025-31883-5

    work page doi:10.1038/s41598-025-31883-5 2025

  29. [29]

    PPGFlowECG: Latent rectified flow with cross-modal encoding for PPG-guided ECG generation,

    X. Fanget al., “PPGFlowECG: Latent rectified flow with cross-modal encoding for PPG-guided ECG generation,” arXiv:2509.19774, 2025

    arXiv 2025

  30. [30]

    SIGMA-PPG: Statistical-prior informed generative mask- ing architecture for PPG foundation model,

    Z. Guoet al., “SIGMA-PPG: Statistical-prior informed generative mask- ing architecture for PPG foundation model,” arXiv:2601.21031, 2026. 11

    arXiv 2026

  31. [31]

    Biometric authentication via hybrid cPPG–rPPG training,

    Z. Sunet al., “Biometric authentication via hybrid cPPG–rPPG training,” inProc. IEEE IJCB, 2024, arXiv:2407.04127

    arXiv 2024

  32. [32]

    Contactless blood pressure measurement via remote photoplethysmography with synthetic data generation using generative adversarial networks,

    B. F. Wuet al., “Contactless blood pressure measurement via remote photoplethysmography with synthetic data generation using generative adversarial networks,”IEEE J. Biomed. Health Inform., vol. 28, no. 2, pp. 621–632, Feb. 2024, doi: 10.1109/JBHI.2023.3265857

    work page doi:10.1109/jbhi.2023.3265857 2024

  33. [33]

    Toward multimodal image-to-image translation,

    J.-Y . Zhuet al., “Toward multimodal image-to-image translation,” in Adv. Neural Inf. Process. Syst. (NeurIPS), 2017, arXiv:1711.11586

    Pith/arXiv arXiv 2017

  34. [34]

    VICReg: Variance-invariance- covariance regularization for self-supervised learning,

    A. Bardes, J. Ponce, and Y . LeCun, “VICReg: Variance-invariance- covariance regularization for self-supervised learning,” inProc. ICLR, 2022

    2022

  35. [35]

    Feasibility of remote pulse transit time estimation using narrow-band multi-wavelength camera pho- toplethysmography,

    G. Slapni ˇcar, W. Wang, and M. Lu ˇstek, “Feasibility of remote pulse transit time estimation using narrow-band multi-wavelength camera pho- toplethysmography,” inProc. ACM Int. Joint Conf. Pervasive Ubiquitous Comput. (UbiComp), 2022, doi: 10.1145/3544793.3560339

    work page doi:10.1145/3544793.3560339 2022

  36. [36]

    Contactless video-based photoplethysmography tech- nique comparison investigating pulse transit time estimation of arterial blood pressure,

    F. Shirbaniet al., “Contactless video-based photoplethysmography tech- nique comparison investigating pulse transit time estimation of arterial blood pressure,” inProc. IEEE Eng. Med. Biol. Soc. (EMBC), 2021, doi: 10.1109/EMBC46164.2021.9629489

    work page doi:10.1109/embc46164.2021.9629489 2021

  37. [37]

    Remote physiological monitoring of neck blood vessels with a high-speed camera,

    M. Cao, G. Saiko, and A. Douplik, “Remote physiological monitoring of neck blood vessels with a high-speed camera,”Adv. Opt. Technol., vol. 14, 2025, doi: 10.3389/aot.2025.1536415

    work page doi:10.3389/aot.2025.1536415 2025

  38. [38]

    High-fidelity rPPG waveform reconstruction from palm videos using GANs,

    T. Li and Y . Liu, “High-fidelity rPPG waveform reconstruction from palm videos using GANs,”Sensors, vol. 26, no. 2, art. 563, 2026, doi: 10.3390/s26020563

    work page doi:10.3390/s26020563 2026

  39. [39]

    How suboptimal is training rPPG models with videos and targets from different body sites?,

    M. Braun, G. Bauer, and M. Elgendi, “How suboptimal is training rPPG models with videos and targets from different body sites?,” inProc. CVPR Workshop, 2024, arXiv:2403.10582

    arXiv 2024

  40. [40]

    Exploiting multiwavelength morphological features of camera-PPG for blood pressure estimation,

    Z. Houet al., “Exploiting multiwavelength morphological features of camera-PPG for blood pressure estimation,”IEEE Trans. Instrum. Meas., 2025, doi: 10.1109/TIM.2025.3551005

    work page doi:10.1109/tim.2025.3551005 2025

  41. [41]

    Depth penetration of light into skin as a function of wavelength from 200 to 1000 nm,

    L. Finlayson, I. R. M. Barnard, L. McMillan, S. H. Ibbotson, C. T. A. Brown, E. Eadie, and K. Wood, “Depth penetration of light into skin as a function of wavelength from 200 to 1000 nm,”Photochem. Pho- tobiol., vol. 98, no. 4, pp. 974–981, Jul. 2022, doi: 10.1111/php.13550

    work page doi:10.1111/php.13550 2022