Organizations: VNU University of Engineering and Technology, Hanoi, Vietnam · College of Engineering and Computer Science, VinUniversity, Hanoi, Vietnam · VinUni-Illinois Smart Health Center, VinUniversity, Hanoi, Vietnam · Center for Innovations in Health Sciences, VinUniversity, Hanoi, Vietnam
Remote photoplethysmography (rPPG) enables non-contact physiological measurement from facial videos; however, its practical deployment is often hindered by substantial performance degradation under domain shift. While recent deep learning-based rPPG methods have achieved strong performance on individual datasets, they frequently overfit to appearance-related factors, such as illumination, camera characteristics, and color response, that vary significantly across domains. To address this limitation, we introduce frequency domain adaptation (FDA) as a principled strategy for modeling appearance variation in rPPG. By transferring low-frequency spectral components that encode domain-dependent appearance characteristics, FDA encourages rPPG models to learn invariance to appearance variations while retaining cardiac-induced signals. To further support physiologically consistent alignment under such appearance variation, we propose Harmonic-Constrained Optimal Transport (HOT), which leverages the harmonic property of cardiac signals to guide alignment between original and FDA-transferred representations. Extensive cross-dataset experiments demonstrate that the proposed FDA and HOT framework effectively enhances the robustness and generalization of rPPG models across diverse datasets.
Figures & tables
Figure 1: Comparison with existing rPPG methods. Conventional approaches are trained on labeled source-domain data and directly deployed to the target domain, resulting in performance degradation under domain shift. In contrast, ours leverages unlabeled target-domain samples to synthesize source data with target-domain appearance during training, encouraging appearance-invariant learning and improving cross-domain generalization.
Figure 2: Overall architecture of HOT.
Figure 3: Illustration of FDA: low-frequency amplitude replacement in the Fourier domain while preserving the phase spectrum.
Model
Venue
PURE → UBFC-rPPG
UBFC-rPPG → PURE
PURE → MMPD
UBFC-rPPG → MMPD
MAE ↓
MAPE ↓
RMSE ↓
r↑
MAE ↓
MAPE ↓
RMSE ↓
r↑
MAE ↓
MAPE ↓
RMSE ↓
r↑
MAE ↓
MAPE ↓
RMSE ↓
r↑
DeepPhys [ 2 ]
ECCV’18
1.46
1.65
3.25
0.98
7.63
7.35
21.97
0.67
22.09
24.38
30.38
0.04
24.50
27.29
31.79
0.01
+ HOT
–
1.31
1.48
3.04
0.99
7.46
7.48
21.95
0.67
19.73
20.07
26.46
0.08
20.65
22.84
24.97
0.05
PhysNet [ 29 ]
BMVC’19
1.70
1.23
3.63
0.99
7.87
15.89
18.29
0.81
16.06
18.46
23.59
0.10
11.25
18.00
23.59
0.34
+ HOT
–
1.17
1.26
3.10
0.99
5.38
12.61
16.52
0.87
12.87
15.90
19.79
0.16
10.36
14.08
18.42
0.38
TSCAN [ 13 ]
NeurIPS’20
1.32
1.49
3.01
0.98
5.98
6.65
19.68
0.73
18.55
20.20
27.13
0.06
18.13
19.92
26.78
0.05
Table 1: Cross-dataset test results. PURE → UBFC-rPPG means the model is trained on PURE and tested on UBFC-rPPG. Lower is better for MAE/MAPE/RMSE, while higher is better for Pearson correlation r .
Figure 4: Cross-dataset evaluation on UBFC-rPPG → MMPD. From left to right: Bland–Altman plot of PhysNet + HOT, Bland–Altman plot of PhysNet, scatter plot of PhysNet + HOT, and scatter plot of PhysNet. The dashed horizontal lines in the Bland–Altman plots denote the 95% limits of agreement, while the dashed diagonal lines in the scatter plots denote the identity line.
Figure 5: Effect of the low-frequency replacement ratio β on cross-dataset performance.
Figure 6: Visual examples of low-frequency replacement under different ratios β .
Figure 7: Performance comparison on PhysNet under progressive integration of the geometric OT loss and the harmonic-constrained OT loss into the baseline.
Figure 8: Heatmap of MAE across different λh values and Sinkhorn iterations. Each cell corresponds to the MAE obtained under a specific hyperparameter combination. Lighter cells indicate lower error (better performance), and the red marker denotes the optimal setting.
Method
Train Time ↓
Peak Memory ↓
Latency ↓
Throughput ↑
(ms/iter)
(GB)
(ms)
(Kfps)
PhysNet w/o HOT
391.3
1.27
1.92
66.75
PhysNet w/ HOT
434.8
2.52
1.89
67.82
Relative Change
+11.1%
+98.4%
-1.6%
+1.6%
Table 2: Computational overhead of HOT on PhysNet [ 29 ] . Training efficiency is measured by per-iteration training time and peak GPU memory. Inference efficiency is measured by forward-only latency and throughput. Lower is better for Train Time, Peak Memory, and Latency; higher is better for Throughput.
Remote photoplethysmography (rPPG) enables non-contact physiological measurement but remains vulnerable to domain shifts from illumination, motion, and sensors. We propose FLOW (Feature-Level Optimal Warping), an optimal transport--driven framework for domain-generalized rPPG. FLOW integrates a Temporal Refinement Module (TRM) to stabilize temporal dynamics and a Prototype-based Cross-Temporal Optimal Transport (PCOT) module to achieve domain-invariant alignment via learnable prototypes.Beyond feature alignment, FLOW employs soft cross-temporal correspondence modeling that aligns temporal features in a flexible manner, allowing the model to respect and preserve the intrinsic rhythmic patterns of physiological signals. Moreover, the lightweight design of our modules allows seamless integration into existing end-to-end rPPG architectures without additional preprocessing. Two regularization terms further enforce source consistency and identity preservation. Theoretically, we derive a generalization bound under conditional optimal transport. Extensive experiments across four rPPG benchmarks show that FLOW achieves state-of-the-art cross-domain performance with lightweight design and strong physiological fidelity.
Bo Zhao, Junzhe Cao, Dan Guo +5
Great Bay University · Harbin Institute of Technology, Shenzhen · Hefei University of Technology +3
Remote photoplethysmography (rPPG) enables contactless physiological monitoring by capturing subtle skin-color variations from facial videos. However, most existing methods predominantly rely on time-domain modeling, making them vulnerable to motion artifacts and illumination fluctuations, where weak physiological clues are easily overwhelmed by noise. To address these challenges, we propose FreqPhys, a frequency-guided rPPG framework that explicitly leverages physiological frequency priors for robust signal recovery. Specifically, FreqPhys first applies a Physiological Bandpass Filtering module to suppress out-of-band interference, and then performs Physiological Spectrum Modulation together with adaptive spectral selection to emphasize pulse-related frequency components while suppress residual in-band noise. A Cross-domain Representation Learning module further fuses these spectral priors with deep time-domain features to capture informative spatial--temporal dependencies. Finally, a frequency-aware conditional diffusion process progressively reconstructs high-fidelity rPPG signals. Extensive experiments on six benchmarks demonstrate that FreqPhys yields significant improvements over state-of-the-art approaches, particularly under challenging motion conditions. It highlights the importance of explicitly modeling physiological frequency priors. The source code will be released.
Wei Qian, Dan Guo, Jinxing Zhou +3
Hefei University of Technology · Hefei Comprehensive National Science Center · MBZUAI +2
Remote photoplethysmography (rPPG) enables contactless measurement of heart rate and other vital signs by analyzing subtle color variations in facial skin induced by cardiac pulsation. Current rPPG methods are mainly based on either end-to-end modeling from raw videos or intermediate spatial-temporal map (STMap) representations. The former preserves complete spatiotemporal information and can capture subtle heartbeat-related signals, but it also introduces substantial noise from motion artifacts and illumination variations. The latter stacks the temporal color changes of multiple facial regions of interest into compact two-dimensional representations, significantly reducing data volume and computational complexity, although some high-frequency details may be lost. To effectively integrate the mutual strengths, we propose PhysNeXt, a dual-input deep learning framework that jointly exploits video frames and STMap representations. By incorporating a spatio-temporal difference modeling unit, a cross-modal interaction module, and a structured attention-based decoder, PhysNeXt collaboratively enhances the robustness of pulse signal extraction. Experimental results demonstrate that PhysNeXt achieves more stable and fine-grained rPPG signal recovery under challenging conditions, validating the effectiveness of joint modeling of video and STMap representations. The codes will be released.
Junzhe Cao, Bo Zhao, Zhiyi Niu +5
Harbin Institute of Technology, Shenzhen · Great Bay University · Hefei University of Technology +1