acceptodds
Under review as a conference paper at ICLR 2027

Reading the Pulse Across Scales: Adaptive Wavelet Decomposition and Interaction

Abstract

Remote photoplethysmography (rPPG) recovers physiological signals from facial videos, where subtle pulse responses are entangled with motion, illumination variations, and imaging noise. These mixed dynamics evolve over different temporal scales and vary across recordings. Existing methods primarily learn from mixed spatiotemporal representations, while predefined wavelet bases offer limited flexibility to accommodate diverse physiological rhythms and interference patterns. To read the pulse across scales, we propose PhysWave, a framework that couples adaptive wavelet decomposition with hierarchical component interaction. PhysWave adaptively combines multiple learnable wavelet experts at each decomposition level to organize local video dynamics into components at different temporal scales. Since temporal decomposition alone does not isolate physiological information, we model intra-component dependencies through sparse attention with a shared, orthogonally initialized projection and learnable gating to refine query–key matching. Temporal spectra then guide the exchange of spatiotemporal features across components, integrating complementary cues across scales for waveform estimation. Experiments on five public rPPG benchmarks demonstrate competitive performance under intra-dataset and cross-dataset evaluation, while ablation studies support the effectiveness of adaptive decomposition and component interaction. The code is available at the \url{https://anonymous.4open.science/r/JSNISE-BEBB.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
rPPG, physiological measurement

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_gN6nOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes PhysWave for recovering rPPG signals from facial videos. AMWD combines multiple learnable wavelet filters according to the input and decomposition level, producing five streams comprising the original features and decomposed components. SOCA models dependencies within each stream, while TSCI uses temporal spectral similarity to guide cross-component feature exchange before waveform fusion.

The authors evaluate the method on five benchmarks through intra-dataset experiments and single-source and multi-source transfer, with VIPL-HR additionally used as a training source. The paper reports competitive heart-rate estimation results, particularly for transfer to MR-NIRP-Car, and includes component ablations and qualitative analysis.

Critical strengths
  1. The paper proposes an adaptive multi-expert wavelet decomposition that allows temporal filtering to vary across recordings. By conditioning the mixture of learnable filters on the input and decomposition level, AMWD provides greater flexibility than applying a fixed wavelet basis to every video. On MMPD, the adaptive design reduces MAE from 3.04 bpm for the best fixed-wavelet alternative to 2.81 bpm. This supports the usefulness of the complete adaptive decomposition module within the proposed architecture.
  2. The paper proposes a framework that combines intra-component refinement with spectrum-guided cross-component interaction. SOCA refines the spatiotemporal features within each component, and TSCI exchanges information according to their spectral relationships, providing a concrete mechanism for combining cues across scales. Across three separate source-to-MR-NIRP-Car transfer settings, the complete model reports average MAE and RMSE of 5.96 and 9.68 bpm, improving over the strongest competing averages in Table 3 by 31.7% and 24.8%, respectively. These results suggest practical value under changes in acquisition conditions.
Critical weaknesses
  1. The novelty needs more precise positioning relative to closely related methods. rPPG-TFCL already incorporates wavelet decomposition into hierarchical rPPG attention. More directly, SOCA’s orthogonally initialized projection and head-wise scaling are closely related to SOatten, while TSCI’s FFT-magnitude descriptors and scaled spectral queries and keys are closely related to FSatten. These mechanisms were introduced in Revisiting Attention for Multivariate Time Series Forecasting (AAAI 2025), which is not discussed in the paper. The component-level application, gating, and normalization may be useful adaptations, but the paper should explain their additional contribution more clearly. AMWD offers a clearer potential distinction, but its comparison with fixed wavelets changes filter learnability, expert count, and input-dependent routing together. The current ablations therefore leave uncertainty about which choices drive the improvement. TSCI’s MAE ablation gain is also small, and variation across runs would help establish its reliability. These limitations reduce confidence in the contribution of individual mechanisms, although the complete framework shows encouraging results.
  2. The training loss in the released code differs from the description in the paper. In PhysWave_Loss, the differentiable loss_distribution_kl returned by the frequency-loss function is not included in the total loss. The included HR_loss instead constructs distributions from detached predictions through NumPy operations, so it provides no gradient to the network. Consequently, the distribution-consistency term does not directly regularize training in this implementation. The authors should confirm whether this code produced the reported results and explain which objective was actually optimized. This matters for reproducing the experiments and interpreting the claimed training constraint.
Questions for authors
  1. What are the substantive technical differences between SOCA/TSCI and SOatten/FSatten? Which differences provide benefits beyond adapting these mechanisms to rPPG components?
  2. With the remaining architecture held fixed, how does AMWD compare with a learnable single filter pair, a learnable four-expert bank with input-independent mixing, and an input-routed bank with random initialization? Please also report variation across runs for the main ablations, particularly the small TSCI gain.
  3. Does the released PhysWave_Loss correspond to the reported experiments? How should the unused differentiable KL term and the detached HR_loss be understood? Please specify the objective used to produce the reported results and whether resolving this discrepancy changes the main findings.
3 comments
AuthorsOpenReview ↗

Response (Part 1)

Thank you for your positive assessment and valuable suggestions. Your professional insights have helped us gain a deeper understanding of our work and sharpen its positioning. For the question that overlaps with those of Reviewer XLCD (Q2), we point to the detailed responses posted under Reviewer XLCD's review and summarize the conclusions.

W1&Q1. Substantive technical differences between SOCA and SOatten, and which differences provide benefits beyond adapting SOatten to rPPG components.

R1. Thank you for this professional suggestion. In essence, SOCA and SOatten [1] share a related idea of orthogonal projection, and TSCI and FSatten [1] share the idea of measuring relations through frequency spectra. However, our modules go well beyond these shared ideas. Our main contribution is a component-centric formulation for weak, interference-dominated rPPG signals: an input-adaptive multi-scale decomposition of facial-video dynamics, followed by refinement within each decomposed component (SOCA) and spectrum-guided exchange across components (TSCI). Within this formulation, SOCA and TSCI contain the following rPPG-specific designs, which we will highlight and strengthen in the revision, together with a discussion of [1].

SOCA. SOCA builds the orthogonal projection into a three-step design that lets the projection play a larger role:

  • Before the orthogonal projection: Queries and keys are generated by temporal center-difference convolution (TDC), which emphasizes the subtle frame-to-frame variations of the pulse, and region-level top-k routing is computed on these original TDC queries and keys, so that fine-grained attention focuses on the regions with the highest query–key similarity.
  • Orthogonal projection: A shared, orthogonally initialized projection with head-wise scaling maps the queries and keys onto new representation directions, providing an alternative feature space that facilitates the separation of coupled channel responses.
  • After the orthogonal projection: A zero-initialized learnable gate adaptively blends the original and projected queries and keys, Q~=Q+tanh(γ)⋅(P−Q)Q̃ = Q + tanh(γ)·(P − Q) with γ = 0 at initialization, so that attention starts from the TDC queries and keys and learns how much projected information to inject, preserving the temporal-difference cues that carry the pulse.

Table S8. Ablation of the three SOCA steps (MAE / RMSE / r). Pre-orth.: top-k routing on the TDC queries and keys; Orth. init. : orthogonally initialized projection; Post-orth.: learnable gate.

VariantPre-orth.Orth. init.Post-orth.MMPD<br>(MAE/RMSE/r)MR-NIRP-Car<br>(MAE/RMSE/r)
w/o SOCA (Table 4)❌❌❌4.01 / 8.23 / 0.8112.23 / 14.47 / 0.05
Orthogonal projection only❌✓❌3.31 / 7.33 / 0.847.94 / 10.02 / 0.20
w/o orthogonal init. (Table 8)†✓❌✓3.00 / 7.16 / 0.847.51 / 8.98 / 0.25
w/o gate✓✓❌3.14 / 6.68 / 0.877.16 / 9.14 / 0.23
Full SOCA✓✓✓2.81 / 6.33 / 0.887.10 / 8.59 / 0.39

†The projection is kept but randomly initialized.

Each step makes the orthogonal projection more effective: the post-orthogonal gate reduces MAE on MMPD from 3.14 to 2.81 and raises r on real-world driving videos (MR-NIRP-Car) from 0.23 to 0.39; orthogonal initialization reduces MAE from 3.00 to 2.81 ( Table 8); and adding the pre-orthogonal routing to the projection changes MAE from 3.31 to 3.14. Overall, the full design reduces MAE from 4.01 without SOCA to 2.81.

[1] Revisiting Attention for Multivariate Time Series Forecasting, AAAI 2025.

AuthorsOpenReview ↗

Response (Part 2)

TSCI. Beyond deriving queries and keys from frequency spectra, TSCI contains two rPPG-specific designs that make cross-component exchange effective:

  • Spectral Normalization: Before the affinity is computed, temporal centering removes the DC term (skin tone and illumination baseline), log compression suppresses dominant motion peaks, per-channel spectral-mean normalization equalizes the energy scales of the components, and L2 normalization of the queries and keys turns the affinity into a comparison of spectral shapes rather than of energy.
  • Diagonal self-retention bias: A learnable non-negative bias softplus(δ)⋅Isoftplus(δ)·I is added to the affinity, preserving each component's sub-band identity during exchange.

In addition, a single affinity per clip and head is shared over all spatiotemporal positions and applied through a zero-initialized gated residual, so the exchange starts from the identity mapping.

Table S9. Ablation of the rPPG-specific designs in TSCI on MMPD. Spec. norm: per-channel spectral-mean normalization; Q/K norm: L2 normalization of queries and keys; Diag. bias: diagonal self-retention bias.

VariantSpec. normQ/K normDiag. biasMAE↓RMSE↓r↑
w/o TSCI–––2.886.430.88
No normalization❌❌✓3.146.800.86
Q/K norm only❌✓✓3.618.170.80
Spec. norm only✓❌✓3.468.100.79
No diag. bias✓✓❌3.989.050.77
Full TSCI✓✓✓2.816.330.88

Every variant that lacks either normalization or the diagonal bias is worse than removing TSCI entirely (MAE 3.14–3.98 vs. 2.88), whereas Full TSCI improves to 2.81. These designs are therefore what make spectral exchange beneficial for rPPG components. The two normalizations are complementary: using only one is even worse than using neither, because the remaining nuisance factor (inter-component energy scale or descriptor norm) then dominates the affinity. Adding the diagonal bias reduces MAE/RMSE by 29.4%/30.1%.

In summary, what PhysWave contributes beyond spectrum- or orthogonality-based attention is (i) the pre-orthogonal routing and post-orthogonal gating that make the orthogonal projection effective in SOCA, and (ii) the spectral normalizations and the diagonal self-retention bias that make spectral exchange beneficial in TSCI. Built on the input-adaptive decomposition, they allow PhysWave to refine each scale and exchange information across scales without erasing the sub-band structure of the decomposition. Thank you again for this suggestion; we will highlight these points in the revised version.

AuthorsOpenReview ↗

Response (Part 3)

W1&Q2. AMWD compared with a learnable single filter pair, a learnable four-expert bank with input-independent mixing, and an input-routed bank with random initialization; relation to rPPG-TFCL; variation across runs.

R2. For the complete control table (Table S1) and the discussion of rPPG-TFCL, please see our response to Reviewer XLCD, W1&Q1 (a) → R1; for the per-seed results (Table S6) and the cross-dataset TSCI ablation (Table S7), please see W2&Q2 (b) → R4 and W2 (c) → R5 there.

We ran all three requested controls with the remaining architecture fixed. In summary:

  • Learnable single filter pair: The best learnable single pair (db6) reaches MAE/RMSE of 2.97/6.41, but the four learnable single experts span MAE 2.97–3.77, so a single basis remains fragile.
  • Learnable four-expert bank with input-independent mixing: It reaches 3.34/8.07 with uniform weights and 2.98/6.75 with globally learned weights, i.e., no better than the best single pair; expert diversity alone does not help.
  • Input-routed bank with random initialization: It reaches 3.66/8.26, versus 2.81/6.33 for AMWD; wavelet initialization reduces MAE/RMSE by 23.2%/23.4%.
  • Relation to rPPG-TFCL: rPPG-TFCL uses a single fixed wavelet basis for every video, whereas AMWD uses four learnable, wavelet-initialized experts with input-dependent routing; PhysWave attains lower RMSE than rPPG-TFCL in all four cross-dataset settings of Table 2.
  • Variation across runs: Over three paired seeds on UBFC-rPPG, the full model outperforms w/o AMWD, w/o SOCA, and w/o TSCI in all 9 (variant, seed) pairs on both MAE and RMSE; without TSCI, MAE/RMSE rise from 0.45 ± 0.05/0.67 ± 0.04 to 0.53 ± 0.02/0.84 ± 0.05, and the TSCI gain is larger under domain shift (RMSE −10.7% to −28.6%).

Therefore, input-dependent routing over a wavelet-initialized expert bank is the key factor behind AMWD, and the component-level improvements are consistent across runs.

W2&Q3. Clarifying the use of the loss function in the released PhysWave code.

R3. Thank you for your attention to the implementation details and for your rigorous attitude. Our contribution lies in the PhysWave architecture (AMWD, SOCA, and TSCI). For the loss function, we directly use the publicly available implementation of RhythmFormer [2] without modification, so that the comparison with this closely related baseline reflects the architectural difference. The loss computed in the code is:

L = 0.2 · L_time + 1.0 · L_CE + 1.0 · L_HR

where L_time is the negative Pearson loss, LCEL_{\mathrm{CE}} is the frequency-domain cross-entropy loss, and LHRL_{\mathrm{HR}} is a heart-rate distribution term.

  • Role of L_HR: As in RhythmFormer [2] , whose Sec. 3.4 describes heart-rate constraints used "to select checkpoints", LHRL_{\mathrm{HR}} is computed from detached predictions and serves as a heart-rate constraint for checkpoint selection on the validation split. It does not provide gradients. The network parameters are therefore optimized by:
L_opt = 0.2 · L_time + 1.0 · L_CE
  • KL term: The loss_distribution_kl value returned by the frequency loss is not included in the training objective.
  • Impact on the main findings: All reported results were produced with this implementation, and all PhysWave variants and ablations use the same objective, so the main findings are unchanged.
  • We have clarified the training objective and the role of LHRL_{\mathrm{HR}} in Sec. 4.1, and we will carefully organize and document the loss code in the released repository.

We thank you again for the constructive feedback. If our responses resolve your concerns, we would be grateful if you would consider further strengthening your support; we are happy to answer any additional questions.

[2] RhythmFormer: Extracting Patterned rPPG Signals Based on Periodic Sparse Attention, Pattern Recognition 2025

Reviewer_SYRfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PhysWave for remote photoplethysmography from facial videos. Starting from a Fusion Stem, it combines four learnable, wavelet-initialized filter experts with sample-dependent routing at each temporal decomposition level. The resulting detail and approximation streams, together with the original stem features, are refined using sparse attention with gated, orthogonally initialized query/key projections. A temporal-spectral interaction module then uses spectral similarities to exchange information across streams before waveform fusion. Evaluation includes intra-dataset experiments on five datasets, single-source transfer experiments, and additional multi-source transfer results; VIPL-HR is used as an additional training source. The reported results are promising, particularly on BUAA and transfer to MR-NIRP-Car, but the evaluation protocol and attribution of the gains need stronger support.

Critical strengths

(1) The reported evaluation covers meaningful acquisition shifts. Beyond controlled intra-dataset tests, the paper includes single-source and multi-source transfer and evaluation on MR-NIRP-Car. In Table 3, PhysWave achieves average MAE/RMSE of 5.96/9.68 bpm, compared with 8.73/12.87 for EfficientPhys, the strongest competing average on these error metrics. If obtained under matched protocols, this is consequential evidence of practical robustness. (2) The decomposition is followed by explicit component interpretation rather than assuming that frequency scale isolates pulse signals. The motivation acknowledges that physiological responses and interference can coexist within each component. Tables 4–6 and 10 provide useful initial evidence for the architecture, including the wavelet alternatives and orthogonal initialization. This makes the proposed combination a plausible direction for learning from mixed physiological and nuisance dynamics.

Critical weaknesses

(1)Important evaluation details remain ambiguous, limiting confidence in the large gains and their comparability. Appendix B explicitly specifies subject-based splits for UBFC-rPPG and MMPD, but describes PURE, BUAA, and MR-NIRP-Car using sequential percentages without identifying the splitting unit or subject IDs. It is therefore unclear whether those partitions are subject-disjoint. A reporting inconsistency also appears for the training-free GREEN method on UBFC-rPPG. Table 2 reports MAE/RMSE of 19.73/21.00, whereas Table 9 reports 19.73/31.00. This discrepancy requires clarification or correction, depending on whether different test populations were used. Until the evaluation protocols and result provenance are explicit, the magnitude of the improvements is difficult to assess. (2)The ablations do not adequately isolate the adaptive-wavelet contribution. Table 5 compares the adaptive expert bank with individual fixed wavelets, leaving unclear whether the gains come from sample-dependent routing, multiple experts, or learnable filters. Matched controls using the same expert bank with uniform or globally learned weights, together with generic-filter initialization, would clarify these contributions. Since the filters are interpolated and trained without explicit wavelet constraints, their learned frequency responses are also needed to support the scale interpretation. Finally, the small TSCI gain (MAE 2.88 to 2.81) is difficult to assess without repeated-run variability.

Questions for authors

(1) Could you clarify the splitting unit and provide the subject IDs for PURE, BUAA, and MR-NIRP-Car, confirming whether the training, validation, and test sets are subject-disjoint? Please also explain why GREEN’s RMSE on UBFC-rPPG differs between Tables 2 and 9 (21.00 versus 31.00), and whether the same test samples and evaluation settings were used. (2) With the expert bank and downstream network held fixed, how does sample-dependent routing compare with uniform or globally learned, sample-independent weights? How does wavelet initialization compare with generic-filter initialization, and do the learned frequency responses retain the intended scale characteristics? Please also provide repeated-run variability to establish whether the TSCI improvement from MAE 2.88 to 2.81 is consistent.

2 comments
AuthorsOpenReview ↗

Response (Part 1)

Thanks for your detailed review and constructive questions, and for noting that our evaluation covers meaningful acquisition shifts and that PhysWave interprets the decomposed components instead of assuming that frequency scale alone isolates the pulse. Several of your questions overlap with those of Reviewer XLCD. For these, we summarize the conclusions here and point to the detailed responses posted under Reviewer XLCD's review.

W1&Q1 (a). The splitting unit and subject IDs for PURE, BUAA, and MR-NIRP-Car, and whether the splits are subject-disjoint.

R1. For the complete subject-ID table (Table S4), the evaluation pipeline, and the provenance of baseline results, please see our response to Reviewer XLCD, W2&Q2 (a) → R3; for a leakage-free check in which BUAA is an unseen target domain, please see W2&Q2 (b) → R4 there.

All dataset splits, preprocessing, and heart-rate evaluation strictly follow the widely used benchmark protocols implemented in the public rPPG-Toolbox [1], and the revised Appendix B now lists the splitting unit and the subject IDs of all five datasets. In summary:

  • All partitions are made at the subject level, and the training, validation, and test subjects are disjoint.
  • All learning-based methods share the same splits, preprocessing, and heart-rate estimation.
  • Test subjects are never used for model selection.
  • PURE uses subjects 1–6 for training and 7–10 for testing; BUAA uses 1–9 for training and 10–13 for testing; MR-NIRP-Car uses 1–12 and 16 for training, 13–14 for validation, and 15 and 17–19 for testing.

Therefore, all three datasets are split in a subject-disjoint manner, and all methods are compared under the same protocol.

W1&Q1 (b). Why GREEN's RMSE on UBFC-rPPG differs between Tables 2 and 9 (21.00 vs. 31.00), and whether the same test samples and settings were used.

R2. Thank you for catching this. This was a transcription error in Table 2, which we have corrected: the correct values are 19.73/31.00/0.37 (MAE/RMSE/r), as in Table 12 and in rPPG-Toolbox [1] (NeurIPS 2023). Same test samples and evaluation settings as in [1].

[1] rPPG-Toolbox: Deep Remote PPG Toolbox, NeurIPS 2023.

AuthorsOpenReview ↗

Response (Part 2)

W2&Q2 (a). With the expert bank and downstream network fixed, how does sample-dependent routing compare with uniform or globally learned weights, and how does wavelet initialization compare with generic-filter initialization?

R3. For the complete controlled ablation (Table R1), including the learnable single-expert variants, please see our response to Reviewer XLCD, W1&Q1 (a) → R1.

With the expert bank and downstream network held fixed, we compared input-dependent (i.e., sample-dependent) routing with uniform and globally learned weights, and wavelet initialization with random (generic) initialization. In summary:

  • Input-dependent routing is the key factor: On MMPD, input-dependent routing achieves MAE/RMSE of 2.81/6.33, versus 2.98/6.75 with globally learned weights (a 5.7%/6.2% reduction) and 3.34/8.07 with uniform weights.
  • Expert diversity alone does not help: The globally weighted bank (MAE 2.98) is no better than the best learnable single expert (2.97), and uniform averaging is even worse than fixed db6 (3.04/6.96).
  • Wavelet initialization matters: Relative to random initialization (3.66/8.26), wavelet initialization reduces MAE/RMSE by 23.2%/23.4%.
  • Filter learning helps, but a single basis remains fragile: Making db6 learnable lowers RMSE from 6.96 to 6.41, but the four learnable single experts span MAE 2.97–3.77.

Therefore, the gain of AMWD comes from adapting the combination of a wavelet-initialized expert bank to each recording, rather than merely from having more or learnable filters.

W2&Q2 (b). Since the filters are interpolated and trained without explicit wavelet constraints, do their learned frequency responses retain the intended scale characteristics and support the scale interpretation?

R4. For the frequency-response analysis of the learned filters (Table S2) and the spectral centroids of the decomposed components (Table S3), please see our response to Reviewer XLCD, W1&Q1 (b) → R2.

Yes. We regard the experts as wavelet-initialized learnable analysis filters (revised Sec. 3.1), and their learned responses support the scale interpretation. In summary:

  • Why no constraint is needed: AMWD is an analysis front-end whose components are never re-synthesized into the input, so perfect reconstruction is not required.
  • The scale hierarchy is fixed by the architecture: The approximation branch is recursively filtered and downsampled by two (Eq. 5), so learning changes how each level splits its band, not the dyadic scale structure.
  • LP/HP roles are preserved: For every expert, the LP centroid remains below 5.0 Hz and the HP centroid above it, and their gap widens slightly after training (on average from 3.17 to 3.35 Hz).
  • Expert complementarity is preserved: The ordering of the HP centroids (sym4 < db4 < db6 < coif3) is unchanged after training.
  • Task-oriented re-tuning rather than collapse: All filters shift only moderately toward lower frequencies (on average −0.59 Hz for LP and −0.41 Hz for HP).
  • Decomposed components remain ordered by scale: The spectral centroids decrease monotonically from D1 to A3.

Therefore, even without explicit wavelet constraints, the learned decomposition retains a meaningful multi-scale structure.

W2&Q2 (c). Repeated-run variability of the TSCI improvement (MAE 2.88 → 2.81).

R5. For the per-seed results of all main ablations (Table S6) and the cross-dataset TSCI ablation (Table S7), please see our responses to Reviewer XLCD, W2&Q2 (b) → R4 and W2 (c) → R5.

In summary, the improvement is consistent. Over three paired seeds on UBFC-rPPG, removing TSCI increases MAE from 0.45 ± 0.05 to 0.53 ± 0.02, and RMSE from 0.67 ± 0.04 to 0.84 ± 0.05, and the full model is better in every run. Under domain shift, TSCI reduces RMSE by 10.7%–28.6% in three cross-dataset settings. Therefore, the TSCI improvement is consistent across runs rather than a run-to-run fluctuation, and it becomes larger under domain shift.

We hope these responses address your concerns about the evaluation protocol and the attribution of the gains. To improve readability, we have also unified the module names and made Appendix B self-contained. If our responses resolve your concerns, we would be grateful if you would consider raising your score; we are happy to discuss further.

Reviewer_XLCDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PhysWave, an rPPG framework that recovers pulse signals from facial videos under motion, illumination changes, and imaging noise. To handle mixed physiological and non-physiological variations across temporal scales, it introduces Adaptive Multi-Expert Wavelet Decomposition for multi-scale representations, Sparse Orthogonal Component Attention for component refinement, and Temporal-Spectral Component Interaction for cross-component exchange. Refined representations predict and fuse candidate rPPG waveforms. Experiments on public benchmarks show competitive performance, especially under challenging illumination and motion, with ablations, visualizations, and cost comparisons.

Critical strengths

1. Strong empirical performance across diverse acquisition conditions. PhysWave is evaluated on controlled, illumination-varied, motion-disturbed, and realistic driving datasets, achieving RMSE of 2.02 bpm on BUAA and 6.33 bpm on MMPD, favorable cross-dataset generalization to MR-NIRP-Car, and additional multi-source appendix evaluations. These results suggest improved rPPG robustness beyond controlled laboratory settings.

2. Well-motivated multi-scale modeling with a compact architecture. PhysWave explicitly constructs multiple temporal representations, refines them individually, and enables spectral-relationship-based exchange, reasonably handling physiological signals and interference with different temporal characteristics. It uses only 0.631M parameters—fewer than all compared models—with competitive computational complexity, and ablations support effectiveness without a large parameter budget.

Critical weaknesses

1. Input-adaptive wavelet decomposition is not sufficiently isolated. AMWD combines learnable wavelet filters, multiple filter experts, and input-dependent routing, but Table 5 does not establish which design choice drives the gain. Missing controls include a learnable single-filter decomposition and a multi-expert filter bank with input-independent routing, leaving unclear whether adaptation to individual recordings or the flexibility of learning and combining multiple filters is responsible. Given prior wavelet-based rPPG work such as rPPG-TFCL, the specific contribution of input-dependent expert combination needs stronger demonstration. The paper also does not specify whether wavelet filter-bank properties are preserved during training, which would clarify the decomposition’s interpretation.

2. Evaluation protocol and reported improvements need clarification. Appendix B describes sequential partitions for PURE, BUAA, and MR-NIRP-Car, but it is unclear whether all partitions are subject-disjoint and whether learning-based baselines used identical splits and preprocessing—important because same-subject recordings share appearance and physiological characteristics. The large BUAA RMSE reduction from 10.17 to 2.02 bpm also lacks random-seed variability or confidence intervals, and repeated-run results would strengthen robustness claims. Evidence for TSCI is limited: removing it changes MAE from 2.81 to 2.88 and RMSE from 6.30 to 6.43, with Pearson correlation unchanged at 0.88; cross-dataset evidence would better establish its practical significance.

Questions for authors

1. Adaptive wavelet routing contribution.

In Table 5, the fixed-wavelet baselines and the full AMWD differ in several respects—learnable filters, multiple experts, and input-dependent routing—so the tested controls do not isolate whether the gain comes from filter learning, expert diversity, or input-dependent adaptation, and it is also unclear whether wavelet filter-bank properties are preserved after initialization; could the authors therefore clarify which of these design choices is responsible for the improvement and whether any filter-bank constraints are enforced during training?

2. Evaluation protocol and result stability.

Appendix B describes sequential partitions for PURE, BUAA, and MR-NIRP-Car, but subject-disjointness and whether learning-based baselines used identical splits, preprocessing, and heart-rate estimation pipelines remain unclear, and the large BUAA RMSE reduction from 10.17 to 2.02 bpm is reported without random-seed variability or confidence intervals; could the authors clarify these protocol details and report multi-seed results to establish the stability of the reported improvements?

5 comments
AuthorsOpenReview ↗

Respinse (Part 1)

Thanks for your careful reading and constructive comments, and for recognizing PhysWave's strong performance across diverse acquisition conditions and its well-motivated, compact design. Here are the detailed responses.

W1&Q1 (a). Which design choice of AMWD (filter learning, expert diversity, or input-dependent routing) drives the gain?

R1. We thank the reviewer for this insightful question. To isolate these factors, we added controlled ablations on MMPD with the downstream network unchanged. As shown in Table S1, the single-expert variants use one wavelet-initialized filter pair, and all bank variants use exactly the same four-expert bank as AMWD, differing only in how the experts are mixed or initialized.

Table S1. Controlled ablation of AMWD on MMPD with the downstream network fixed. Input-independent mixing applies the same expert weights to every input video (fixed uniform weights or globally learned weights); input-dependent mixing predicts the weights from each input video at each decomposition level (i.e., sample-dependent routing).

VariantFilters#ExpertsExpert mixingInit.MAE↓RMSE↓r↑
db6fixed1–wavelet3.046.960.86
sym4fixed1–wavelet3.477.820.81
db4fixed1–wavelet3.487.890.81
coif3fixed1–wavelet3.487.880.81
db6learnable1–wavelet2.976.410.87
coif3learnable1–wavelet3.106.940.86
db4learnable1–wavelet3.407.680.82
sym4learnable1–wavelet3.777.930.82
Uniform weights (0.25 each)learnable4input-independentwavelet3.348.070.80
Globally learnable weightslearnable4input-independentwavelet2.986.750.86
Input-adaptive routinglearnable4input-dependentrandom3.668.260.79
Input-adaptive routing (Ours)learnable4input-dependentwavelet2.816.330.88

These controls lead to four conclusions:

  • Filter learning helps, but a single basis remains fragile: Making db6 learnable lowers RMSE from 6.96 to 6.41, yet the four learnable single experts span MAE 2.97–3.77, so the outcome depends on choosing the right basis in advance.
  • Expert diversity alone does not help: With the same bank, globally learned (sample-independent) weights give MAE 2.98, no better than the best learnable single expert (2.97), and uniform averaging (3.34/8.07) is even worse than fixed db6. A single combination applied to every recording behaves like a single analysis.
  • Input-dependent routing is the key factor: Changing only the mixing from global to input-adaptive (same bank, same network) reduces MAE/RMSE by 5.7%/6.2% and raises r from 0.86 to 0.88. AMWD also outperforms the best learnable single expert without requiring any basis selection. Consistently, Fig. 4 shows that the routing weights shift with the recording condition.
  • Wavelet initialization matters: With the same input-adaptive routing, wavelet initialization reduces MAE/RMSE by 23.2%/23.4% relative to random initialization. Thus, the wavelet prior provides a structured starting point that free filter learning does not recover.

Relation to rPPG-TFCL [1]: rPPG-TFCL embeds a standard 2D DWT with a single predefined wavelet basis for one-level spatial decomposition into a hierarchical spatio-temporal transformer and couples it with time–frequency consistency losses. AMWD differs from it in three key respects, each of which Table S1 shows to matter:

  • One basis vs. four experts: rPPG-TFCL decomposes every video with a single wavelet basis, whereas AMWD uses a bank of four wavelet-initialized experts with complementary frequency characteristics.
  • Fixed vs. learnable filters: The DWT filters in rPPG-TFCL are fixed, whereas the AMWD experts are learnable and re-tuned for the task (R2).
  • Input-independent vs. input-dependent decomposition: rPPG-TFCL applies the same decomposition to every video, whereas AMWD routes the experts according to each input and decomposition level. This is the most important factor in Table S1.

Consistently, our PhysWave attains lower RMSE than rPPG-TFCL in all four cross-dataset settings of Table 2. We have added this discussion to Sec. 2 and the controlled ablation to Sec. 4.4.

[1] rPPG-TFCL: Time-Frequency Consistency Learning for Robust Remote Physiological Measurement, Knowledge-Based Systems 2025.

AuthorsOpenReview ↗

Response (Part 2)

W1&Q1 (b). Are wavelet filter-bank properties preserved, and are any filter-bank constraints enforced during training?

R2. Thank you for raising this point. No explicit filter-bank constraint (e.g., perfect reconstruction) is enforced during training, and this is by design. However, the properties on which the scale interpretation relies are preserved, as verified below.

  • Why no constraint is needed: AMWD is an analysis front-end for a discriminative task. Its components are never re-synthesized into the input, so perfect reconstruction is not required. Therefore, we regard the experts as wavelet-initialized learnable analysis filters, and the revised Sec. 3.1 states this explicitly.
  • The scale hierarchy is fixed by the architecture: The approximation branch is recursively filtered and downsampled by a factor of two (Eq. 5), so learning changes how each level splits its band, not the dyadic scale structure.
  • Learned frequency responses: We measured the frequency response of each expert's low-pass (LP) and high-pass (HP) filter before and after training (fs = 30 Hz). Because the experts are interpolated to a common kernel length (Sec. 3.1), the effective LP/HP crossover of the shorter filters lies below the nominal half-band frequency of 7.5Hz. Therefore, we use 5.0 Hz, which separates the LP and HP centroids of all experts both before and after training, as the observation boundary and report the share of each filter's energy above it (Table S2).

Table S2. Frequency characteristics of the four AMWD experts. Each entry reads before training → after training; "energy > 5 Hz" is the share of a filter's spectral energy above 5.0 Hz.

ExpertLP centroid (Hz)HP centroid (Hz)LP energy > 5Hz (%)HP energy > 5Hz (%)
db44.174 → 3.5706.243 → 5.87827.952 → 22.24457.459 → 53.073
db64.251 → 3.8487.918 → 7.60828.720 → 25.50784.764 → 80.644
sym44.281 → 3.5635.989 → 5.59430.126 → 23.83653.783 → 49.133
coif34.876 → 4.25710.123 → 9.56643.148 → 34.76291.343 → 85.499
Average4.396 → 3.8107.568 → 7.16232.487 → 26.58771.837 → 67.087

The analysis shows three properties:

  • LP/HP roles are preserved: For every expert, the LP centroid remains well below the HP centroid, and the gap even widens slightly after training (on average from 3.17 to 3.35 Hz).
  • Expert complementarity is preserved: The ordering of the HP centroids (sym4 < db4 < db6 < coif3) is unchanged, and coif3 remains by far the most high-frequency-selective expert (HP centroid 9.566 Hz vs. 5.594-7.608 for the others).
  • Task-oriented re-tuning rather than collapse: All filters shift moderately toward lower frequencies (on average −0.59 Hz for LP and −0.41 Hz for HP), and the share of LP energy above 5.0 Hz drops from 32.487% to 26.587. The approximation path, through which the pulse band reaches the coarser levels, therefore suppresses high-frequency noise more strongly.
  • Decomposed components remain ordered by scale: The spectral centroids of the resulting components decrease monotonically from D1 to A3 (Table S3), i.e., the components remain distinct and ordered from high to low relative frequency rather than converging to a common band.

Table S3. Spectral centroids of the decomposed components after training (visualized in the revised Appendix D.1).

ComponentD1D_1D2D_2D3D_3A3A_3
Spectral centroid (Hz)1.9801.4331.1960.983

We have added the before/after frequency-response curves to the revised paper and the component spectra and centroids to Appendix D.1.

AuthorsOpenReview ↗

Response (Part 3)

W2&Q2 (a). Are all partitions subject-disjoint, and did learning-based baselines use identical splits, preprocessing, and heart-rate estimation?

R3. Yes. All dataset splits, preprocessing, and heart-rate evaluation strictly follow the widely used benchmark protocols [2–6], implemented in the public rPPG-Toolbox [2]. All five intra-dataset partitions are made at the subject level, with no subject overlap between the training, validation, and test sets. The subject IDs are listed in Table S4 and have been added to Appendix B.

Table S4. Subject IDs of the intra-dataset splits.

DatasetTrainValidationTest
PURE1–6–7–10
UBFC-rPPG1, 3, 4, 10–18, 20, 22–27, 30–40–5, 8, 9, 41–49
MMPD1–2324–2627–33
BUAA1–9–10–13
MR-NIRP-Car1–12, 1613, 1415, 17–19
  • Splitting unit: Every partition is made over subjects, exactly as listed above. For UBFC-rPPG, the "first 30 subjects" follow the subject-folder ordering of rPPG-Toolbox [2]. Some subject numbers are missing because the original dataset does not contain data for those subjects, hence the non-contiguous IDs.
  • Identical pipeline: All learning-based methods use these splits, the standard rPPG-Toolbox preprocessing, and the same heart-rate estimation (a first-order Butterworth band-pass filter of 0.75–2.5 Hz followed by Welch PSD peak detection), following [2–6]. Baseline results were either reproduced under these splits or taken from works that use the identical protocol. Appendix B now states the source of each baseline result.
  • Model selection: The validation split is used where one exists (MMPD, MR-NIRP-Car), and the last epoch is used otherwise. Test subjects are never used for model selection.

[2] rPPG-Toolbox: Deep Remote PPG Toolbox, NeurIPS 2023.

[3] RhythmFormer: Extracting Patterned rPPG Signals Based on Periodic Sparse Attention, Pattern Recognition 2025.

[4] PhysDiff: Physiology-Based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement, AAAI 2025.

[5] PhysLLM: Harnessing Large Language Models for Cross-Modal Remote Physiological Sensing, ICLR 2026.

[6] Phase-Net: Physics-Grounded Harmonic Attention System for Efficient Remote Photoplethysmography Measurement, CVPR 2026.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all three reviewers for their valuable feedback and constructive comments on our paper. Their insights have been instrumental in significantly strengthening the attribution analysis, evaluation protocol, and experimental rigor of our work. We are also encouraged that the reviewers recognized PhysWave's strong performance across diverse acquisition conditions, including transfer to real-world driving videos (XLCD, SYRf, gN6n); its well-motivated multi-scale design with a compact architecture (XLCD); its input-adaptive decomposition, which offers more flexibility than applying a fixed wavelet basis to every video (gN6n); and its explicit modeling within and across the decomposed components rather than assuming that frequency scale alone isolates the pulse (SYRf, gN6n).

Summary of Addressed Reviewer Concerns

Our revised manuscript systematically addresses all key issues raised by the reviewers through new controlled experiments, quantitative analyses, explicit protocol details, and corrections. Detailed point-by-point responses are posted under each review.

Reviewer XLCD:

We addressed all five concerns raised by Reviewer XLCD by providing controlled ablations, a frequency-response analysis, explicit protocol details, and multi-seed and cross-dataset evidence:

  • Attribution of AMWD: We added controlled ablations that separate filter learning, expert diversity, input-dependent routing, and initialization, which identify input-dependent routing as the key factor (R1).
  • Filter-bank properties: We clarify that no filter-bank constraint is imposed by design and analyze the learned frequency responses, which retain their low-/high-pass roles and the complementary characteristics of the experts (R2).
  • Evaluation protocol: We confirm that all splits are subject-disjoint, list every subject ID, and specify the shared preprocessing and heart-rate estimation pipeline (R3).
  • Stability of the improvements: We added multi-seed results with confidence intervals on BUAA, repeated runs of the main ablations, and a leakage-free check in which BUAA is an unseen target domain (R4).
  • Significance of TSCI: We added cross-dataset ablations, in which the benefit of TSCI is substantially larger than in the intra-dataset setting (R5).

Reviewer SYRf:

We addressed the concerns of Reviewer SYRf regarding the evaluation protocol and the isolation of the adaptive-wavelet contribution:

  • Splitting unit and result provenance: We provide the subject IDs and confirm subject-disjointness (R1). The GREEN discrepancy was a transcription error, which we corrected after auditing (R2).
  • Isolating the adaptive-wavelet contribution: We added matched controls with uniform and globally learned weights and with random initialization (R3), analyzed the learned frequency responses (R4), and reported the repeated-run variability of TSCI (R5).

Reviewer gN6n:

We addressed the questions of Reviewer gN6n regarding positioning, the AMWD controls, and the use of the loss function:

  • Positioning relative to SOatten/FSatten: We now discuss this work, highlight the rPPG-specific designs of SOCA and TSCI, and show by ablation that these designs are necessary and effective (R1).
  • AMWD controls, relation to rPPG-TFCL, and run-to-run variation: All requested controls and multi-seed ablations are provided (R2).
  • Use of the loss function in the code: We clarify how the loss function is used in the released code and in Sec. 4.1; all reported results were produced with this implementation, and the main findings are unchanged (R3).

Conclusion

We believe these comprehensive revisions, especially the controlled ablations that isolate the contribution of input-dependent routing, the explicit subject-disjoint protocol, and the multi-seed and cross-dataset evidence, solidly address the reviewers' concerns regarding the attribution of the gains, the evaluation protocol, and the novelty of PhysWave. All new analyses, experimental results, and textual clarifications have been integrated into the revised manuscript and are highlighted in red: the discussion of related work (Sec. 2), the clarified filter design and training objective (Secs. 3.1 and 4.1), the new ablations and frequency-response analysis (Sec. 4.4 and Appendices C and D.1), and the complete splitting protocol (Appendix B). We are happy to answer any further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.