acceptodds
Under review as a conference paper at ICLR 2027

KDPPG: Prior-Guided Knowledge Distillation for Robust Streaming rPPG Estimation

Abstract

Remote photoplethysmography (rPPG) enables contactless physiological monitoring, but robust low-latency deployment remains challenging. High-performance methods typically rely on buffered video clips, whereas streaming-compatible models often provide weaker temporal modeling and robustness. We propose KDPPG, a prior-guided knowledge distillation framework that transfers spatial and geometric knowledge from a clip-wise teacher to a causal student. KDPPG-T uses facial masks and projected normalized coordinate code (PNCC) in a dual-prior refinement module to suppress background illumination artifacts and improve robustness to motion. KDPPG-S processes one frame per time step, maintains a recurrent state, and incorporates a parallel harmonic resonance module (PHRM) to model periodic dynamics without fixed clip buffering or inference-time priors. Extensive experiments across multiple datasets demonstrate that KDPPG achieves state-of-the-art accuracy, while KDPPG-S offers high inference efficiency and low computational cost, making it suitable for practical streaming deployment.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
rPPG, Heart rate estimation, Knowledge distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_W2QHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

KDPPG uses knowledge distillation to train a streaming rPPG model. A clip-based teacher incorporates facial masks and PNCC geometry features to reduce interference from lighting and motion. The student receives supervision from both the teacher’s features and predicted waveforms. It combines Mamba-2 with a parallel harmonic resonance module (PHRM) and processes incoming frames while maintaining temporal states. Experiments cover four datasets, two cross-dataset settings, component ablations, and computational profiling. The reported results show that the student achieves accuracy close to the teacher without requiring its auxiliary priors at inference.

Critical strengths
  1. The teacher combines explicit background and geometry cues to address two relevant sources of interference. DPRM uses a facial mask to separate face and background features, subtracts pooled background features as an interference reference, and incorporates PNCC geometry through cross-attention. These are concrete additions to the teacher design. In Tables 3–4, removing the mask or PNCC increases MMPD MAE from 2.731 bpm to 3.611 and 3.414 bpm, respectively. These results support the usefulness of the priors for improving the teacher and providing a stronger source of supervision.
  2. The student introduces explicit oscillatory dynamics alongside Mamba-2 for periodic signal modeling. PHRM uses learnable damped rotations to update its states and combines its output with the Mamba-2 branch through a learned gate. This gives the recurrent student a direct mechanism for representing physiological rhythms. Table 5 shows a corresponding improvement: MMPD MAE decreases from 3.45 to 2.95 bpm, while correlation increases from 0.79 to 0.86. This is the clearest evidence for a useful architectural contribution in the student.
Critical weaknesses
  1. The contribution beyond KDPhys is not sufficiently established. KDPhys already transfers spatiotemporal representations from a 3D CNN teacher to a lightweight 2D CNN student for rPPG. KDPPG introduces teacher priors, a causal recurrent student, and PHRM, which are potentially meaningful extensions. However, the paper neither discusses this closely related work nor provides comparisons that establish the value of these extensions. The teacher ablations show that priors improve the teacher, while the PHRM ablation shows that the periodic branch improves the student. Neither isolates the benefit of transferring prior-guided representations. Figure 8 also changes teacher architectures rather than isolating the priors. Without a matched student trained without distillation and a controlled comparison of teacher priors, I cannot determine how much of the improvement comes from the proposed distillation framework.
  2. The evidence for low-latency physiological measurement is incomplete. Producing a waveform sample for each frame does not establish how soon a reliable HR estimate becomes available. Appendix D does not specify whether Butterworth filtering is causal or how much signal Welch estimation requires. The efficiency results also compare one-frame and 160-frame forward passes, with clip-based methods repeatedly processing stride-one windows. This is a relevant setting for frequent updates, but the advantage depends on that update policy. The paper needs to establish its accuracy and latency under a complete causal pipeline and a comparable output schedule, since streaming deployment is a central claim.
Questions for authors
  1. Clarification of the contribution: What capability does KDPPG add beyond KDPhys? Please distinguish the contributions of teacher priors, the causal student, and PHRM, and identify which existing experiments support those distinctions.
  2. Controls needed to assess the central claim: Can you report results for the same student trained without distillation, and for distillation from the same teacher with its priors removed? These controls would help determine whether the gains require prior-guided knowledge transfer or mainly follow from the student architecture and teacher quality.
  3. Clarification and validation of streaming evaluation: Were Tables 1–2 obtained through sequential inference with persistent states and past-only processing? Please specify state resets, the filtering implementation, and the HR estimation window. Under that pipeline, how does HR error change with elapsed observation time, and does the efficiency advantage remain at a matched output refresh interval?
4 comments
AuthorsOpenReview ↗

Response to Reviewer W2QH (Part 1/2)

We thank the reviewer for recognizing the value of teacher priors and periodic modeling, and for highlighting the need to distinguish the evidence for knowledge distillation from that for streaming deployment. We respond point by point below.

1. Contributions beyond KDPhys (W1 & Q1)

Beyond KDPhys, KDPPG provides a more general distillation framework and develops a state-of-the-art teacher and a reusable student. KDPhys relies on predefined teacher–student layer correspondences, limiting direct reuse across architectures. Adapting it to a new teacher or student requires reconfiguring the distillation locations and feature alignment when these correspondences no longer hold. Our framework supervises deep features and predicted waveforms without requiring layer-by-layer correspondence, allowing the student architecture to remain fixed when changing teachers. Appendix E.2 demonstrates this capability by using PhysNet, RhythmMamba, and KDPPG-T to supervise the same KDPPG-S student.

We will add the citation and discussion in the revised manuscript. We omitted it from the original comparison because we could not find publicly available official code. We have reimplemented KDPhys based on the details provided in its paper and compared both methods using the same MMPD data splits and HR evaluation protocol.

MethodMAE↓RMSE↓MAPE↓ρ↑
KDPhys-T3.839.233.860.74
KDPPG-T2.736.862.760.85
KDPhys-S4.089.624.040.72
KDPPG-S2.956.833.130.86

The three components contribute in distinct ways (All table and figure numbers refer to the revised manuscript.):

  • We design a dual-prior refinement module (DPRM) in the teacher, combining mask-guided artifact suppression and geometry-guided cross-attention to improve robustness to illumination changes and head motion, as demonstrated by the ablations in Tables 4–5.
  • We design a causal student that retains temporal states across incoming frames, enabling frame-wise waveform inference without clip buffering or inference-time priors (Sec. 3.3). Table 7 reports its computational cost and throughput.
  • We design PHRM within the student to explicitly model periodic physiological dynamics through learnable damped rotations. Fig. 3 visualizes the resulting periodic structure, and Table 6 demonstrates the accuracy improvement from adding PHRM.

2. Contributions of distillation and teacher priors (W1 & Q2)

Distillation improves the student even with a prior-free teacher, while the prior-guided teacher provides further gains. To assess the contribution of prior-guided knowledge transfer, we compare the same student trained without distillation and with distillation from teachers sharing the same RGB backbone, with and without mask and PNCC priors. All student settings retain ground-truth supervision.

ModelTraining / distillation settingMAE↓RMSE↓MAPE↓ρ↑
Teacher1Full priors2.736.862.760.85
Teacher2w/o mask, w/o PNCC3.928.733.980.77
Student0w/o distillation4.029.224.150.73
Student1.1Response-level only (Teacher1)3.317.403.570.83
Student1.2Feature-level only (Teacher1)3.499.053.540.76
Student1.3Response-level + feature-level (Teacher1)2.956.833.130.86
Student2Response-level + feature-level (Teacher2)3.758.873.900.76

Our response continues in the next comment, where we address Question 3 on streaming evaluation, HR observation time, and efficiency at matched output refresh intervals.

AuthorsOpenReview ↗

Response to Reviewer W2QH (Part 2/2)

3. Causal evaluation and HR observation time (W2 & Q3)

For a fair comparison with existing methods, Tables 1–2 follow the full-video HR evaluation protocol used by rPPG-Toolbox. The student resets its states at the beginning of each 128-frame clip, then processes frames sequentially and outputs one rPPG sample per frame. The predicted waveforms are concatenated chronologically, and HR is estimated from the assembled prediction and the reference PPG of the same length using identical postprocessing. Thus, 128 frames is the network processing length, rather than the HR observation window in the main tables. This protocol ensures consistent accuracy evaluation across methods.

To evaluate online startup, when sufficient historical information has not yet accumulated, we further compare HR estimation errors at different observation lengths. KDPPG-S resets its states only at the beginning of the video and subsequently accumulates historical information, while RhythmMamba processes the first N frames observed so far. Both methods estimate HR from the predicted waveform and contact PPG of the corresponding length. HR is estimated using Butterworth bandpass filtering (0.75–2.5 Hz, forward–backward implementation) and Welch estimation, both restricted to the currently observed signal window without accessing future data.

KDPPG-S produces an accurate HR estimate from only 3 seconds of observed video, whereas the state-of-the-art method RhythmMamba cannot yet provide a meaningful HR estimate at this point. These results demonstrate that KDPPG-S reaches stable HR estimation with a shorter observation time.

Observed framesObservation time (s)KDPPG-S MAE↓RhythmMamba MAE↓
602.007.9611.35
702.335.417.42
802.674.4813.28
903.003.4011.50

Once enough video frames have been collected for all methods to produce stable and accurate HR estimates, we compare network throughput at different refresh strides during online inference. KDPPG-S uses a mathematically equivalent parallel implementation of multi-frame recurrence throughout the network, including PHRM. It therefore supports multi-frame input and rPPG output of the corresponding length: at a refresh stride of r frames, KDPPG-S processes only the r newly arrived frames, whereas clip-wise methods reprocess the latest 128 frames. For these throughput measurements, stride-1 inference retains recurrent states, while larger strides reset the states at the beginning of each chunk. All methods are evaluated on the same RTX 4090D GPU using FP32, a batch size of 1, and a spatial resolution of 128×128. We measure network inference time only, excluding initial observation, data transfer, and pre/postprocessing. Throughput is calculated as the number of newly arrived video frames divided by network inference time, in frames per second.

With matched output refresh intervals, KDPPG-S achieves the highest throughput. As the refresh stride increases, the student can process newly arrived frames together, reducing the overhead of per-frame calls. Its throughput advantage therefore extends from frame-wise updates to block execution at lower refresh frequencies.

MethodFrames per callStride 1Stride 15Stride 30Stride 45
KDPPG-SMatches stride: 1 / 15 / 30 / 45430.502832.093059.503127.40
RhythmMamba12849.10736.661473.392275.80
PhysNet12854.82825.131648.522501.63
PhysFormer12840.82612.311224.421854.75

Thank you for the additional experiments. The comparison with KDPhys and the use of multiple teachers with the same student clarify the practical contribution. The controlled teacher ablation addresses my main concern: with the student held fixed, MAE decreases from 4.02 bpm without distillation to 3.75 bpm with a teacher that uses no auxiliary priors, and to 2.95 bpm with the prior-guided teacher. The prefix-only HR evaluation also provides more direct evidence for early HR estimation. One point should be made explicit in the revision: the startup accuracy experiment carries state across the video, whereas the throughput measurements at larger strides reset state between chunks. These results should be presented as measurements under distinct protocols. I am raising my score to 3 (Weak acceptance).

Thank you for recognizing our work and raising the score. We have clarified the different state-reset settings for the startup accuracy and throughput experiments in the captions of Tables 8 and 9 in the revised manuscript.

Reviewer_CrwPOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper aims to reducing the latency of Remote photoplethysmography (rPPG) model while maintaining high accuracy and robustness. To realize this, this paper designs a clip-wise teacher whose both accuracy and latency is high, and a frame-wise student which normally has lower accuracy and lower latency. Via the proposed prior-guided knowledge distillation technique, the frame-wise student distills useful knowledge from the clip-wise teacher. Experiments verify the effectiveness of the proposed method.

Critical strengths
  • The paper is generally easy to follow and understand.

  • The solution is reasonable.

Critical weaknesses
  • Knowledge distillation is a quite popular and common technique. I didn't quite understand the contribution of this paper from the technical view. What is the specific challenge in the context of rPPG? And what is the specific design tailed for the challenge?
Questions for authors
  • As video clip and frame do not contain equal amount of information, why is it possible to distill knowledge from the clip-wise teacher to a frame-wise student?
3 comments
AuthorsOpenReview ↗

Response to Reviewer CrwP

We thank the reviewer for finding the paper clear and the proposed solution reasonable.

1. Technical challenge and our tailored design (W1)

Our contribution is a prior-guided clip-wise-to-frame-wise distillation framework that transfers robust pulse representations to a causal student, enabling streaming rPPG estimation without clip buffering or inference-time priors.

The specific challenge is to recover weak, quasi-periodic pulse signals under illumination changes and head motion while avoiding the acquisition latency of collecting video clips. Frame-wise spatial features and adjacent-frame differences alone provide insufficient temporal context to capture physiological rhythms. We therefore design complementary mechanisms for extracting and transferring this context: the teacher uses DPRM to suppress interference through facial masks and geometry priors, while the student combines persistent historical states with explicit periodic memory through PHRM. Feature-level and response-level distillation train the student to retain pulse-relevant information from sequential inputs using the teacher's cleaner representations and waveform targets.

2. How frame-wise input uses historical information (W1 & Q1)

Our student receives one new frame at a time while retaining the internal states of Mamba-2 and PHRM. For example, the Mamba-2 state update can be simplified as

= A_t A_{t-1} h_{t-2} + A_t B_{t-1} x_{t-1} + B_t x_t,$$ where $x_t$ is the current frame feature and $h_t$ is the temporal state. This expansion shows that the current state contains both the current input and information retained from previous inputs. Thus, receiving one frame at a time does not mean making predictions from that frame alone, historical information continues to contribute through the state. To capture pulse periodicity, PHRM further uses a damped rotational update $$q_t = \lambda R(\omega) q_{t-1} + (1-\lambda)u_t.$$ Here, $R(\omega)$ rotates the periodic state $q_{t-1}$ forward in time, $\lambda$ controls damping, and $u_t$ incorporates new input to refine the state. Mamba-2 accumulates historical features, while PHRM provides explicit memory for periodic dynamics. Together, they support continuous pulse waveform estimation without repeatedly processing a full video clip. ### 3. Why clip-wise-to-frame-wise distillation is feasible (W1 & Q1) **Clip-wise-to-frame-wise distillation is feasible because the student learns to approximate the teacher using the current frame and accumulated historical states, without requiring identical teacher–student inputs.** The teacher uses full clips and auxiliary priors to provide cleaner feature and waveform targets at each time step. Since the student's outputs depend on its historical states, matching these targets also trains its state updates to retain useful pulse cues and suppress interference. The continuity and quasi-periodicity of pulse signals make past observations informative for current predictions, allowing the student to encode the teacher's temporal patterns in recurrent state updates.

Dear Reviewer,

As the ICLR 2027 discussion phase ends on November 18 at 11:59 PM AoE, please take a moment to review and respond to the authors' comments, even if your response is simply an acknowledgement. Your active participation in the discussion is greatly appreciated.

Thank you for your time and effort!

Dear Reviewer CrwP,

With only approximately an half hour remaining before thediscussion period closes, we hope you can understand our eagerness to hear your thoughts.

We would be truly grateful for any further feedback youmay have. Thank you again for your time and efforts inreviewing our work.

We wish you all the best!

Reviewer_18SzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KDPPG for streaming rPPG estimation. A clip-wise teacher uses facial masks and PNCC as priors. A student combines Mamba-2 with a parallel harmonic resonance module and learns from the teacher.

Critical strengths
  1. Transferring knowledge from a prior-guided teacher to a causal student is a well-motivated.
  2. The experiments include four datasets, cross-dataset evaluation, and component ablations. The results are promising
Critical weaknesses

The paper does not clearly show how much knowledge distillation helps. Appendix E compares different teachers, but there is no comparison with the same student trained only using ground-truth labels. Knowledge distillation has already been used for efficient rPPG estimation. The paper should explain how and why mor specifically KDPPG differs from this closely related work and provide a comparison under the same evaluation protocol. The feature-level and response-level distillation losses are also not tested separately. These comparisons would help show where the gains come from.

Questions for authors

The results are promising, but knowledge distillation is the central claim of this paper, and its contribution is not clearly demonstrated. The missing comparison with closely related work and the lack of a student baseline without distillation leave the main source of improvement unclear. I recommend Weak Reject.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 18Sz

We thank the reviewer for recognizing the motivation for transferring knowledge from a prior-guided teacher to a causal student, and the promising results across four datasets, cross-dataset evaluations, and component ablations. We address the contribution of knowledge distillation and the comparison with closely related work below.

1. Contribution of knowledge distillation (W1 & Q1)

Knowledge distillation improves the accuracy of the same student: response-level and feature-level supervision each provide gains, and combining them achieves the best performance. To quantify these contributions, we add controlled comparisons with no distillation, response-level distillation only, feature-level distillation only, and both. All distillation variants use the same KDPPG-T teacher, and all settings retain ground-truth supervision and are evaluated on MMPD under the same protocol.

Distillation settingMAE↓RMSE↓MAPE↓ρ↑
No distillation4.029.224.150.73
Response-level only3.317.403.570.83
Feature-level only3.499.053.540.76
Response-level + feature-level2.956.833.130.86

2. Comparison with KDPhys (W1 & Q1)

Compared with KDPhys, KDPPG allows different teacher architectures to supervise the same efficient frame-wise student without requiring layer-by-layer correspondence. KDPPG also achieves state-of-the-art performance across multiple benchmarks (Tables 1–2).

We will add the citation and discussion in the revised manuscript. We omitted it from the original comparison because we could not find publicly available official code. We have now reimplemented it based on the details provided in its paper. Under the same MMPD data splits and HR evaluation protocol, both our teacher and distilled student achieve better estimation accuracy:

MethodMAE↓RMSE↓MAPE↓ρ↑
KDPhys-T3.839.233.860.74
KDPPG-T2.736.862.760.85
KDPhys-S4.089.624.040.72
KDPPG-S2.956.833.130.86

KDPhys relies on feature alignment across multiple corresponding teacher–student layers, requiring these correspondences to be reconfigured when changing architectures. Our framework instead supervises deep features and predicted waveforms without requiring layer-by-layer correspondence. Appendix E.2 demonstrates this flexibility using PhysNet, RhythmMamba, and KDPPG-T as teachers for the same student. This allows future, more effective rPPG models to serve as teachers without redesigning the student network.

Our teacher incorporates a dual-prior refinement module (DPRM), combining mask-guided artifact suppression and geometry-guided cross-attention to improve robustness to illumination changes and head motion. Our student uses a specially designed temporal modeling module that combines persistent historical memory with explicit periodic memory provided by PHRM. Compared with the local temporal shifts used by KDPhys, this design captures long-range temporal dependencies and sustained physiological rhythms while supporting causal frame-wise inference without clip buffering.

Dear Reviewer 18Sz,

With only approximately an half hour remaining before thediscussion period closes, we hope you can understand our eagerness to hear your thoughts.

We would be truly grateful for any further feedback youmay have. Thank you again for your time and efforts inreviewing our work.

We wish you all the best!

Comments on the paper

1 comment

We thank the reviewers for their helpful feedback and recognition of our work.

KDPPG combines a more general distillation framework, a robust prior-guided teacher, and an efficient causal student.

  • A more general distillation framework. Deep-feature and waveform supervision enable knowledge transfer between teachers and students with different architectures, without requiring similar network structures or layer-by-layer correspondence.
  • A strong prior-guided teacher. Facial masks and geometry priors improve robustness to illumination changes and head motion, enabling state-of-the-art accuracy across multiple benchmarks.
  • An efficient, reusable student. Our temporal modeling design captures long-range dependencies and explicit periodic dynamics through PHRM, supporting causal frame-wise inference without clip buffering. The same student can learn from different teacher architectures.

We appreciate the reviewers’ time and thoughtful suggestions, which have helped us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.