Skip to content

PHASE-Net: Physics-Grounded Harmonic Attention System for Efficient Remote Photoplethysmography Measurement

Conference: CVPR 2026 arXiv: 2509.24850 Code: GitHub Area: Human Understanding Keywords: rPPG, physics-informed network, temporal convolutional network, hemodynamics, Navier-Stokes, lightweight model

TL;DR

Starting from the Navier-Stokes equations, this work derives through rigorous mathematical analysis that rPPG pulse signals obey a second-order damped harmonic oscillator model whose discrete solution is equivalent to a causal convolution operator, thereby providing a first-principles justification for the TCN architecture. The resulting PHASE-Net, with only 0.29M parameters, achieves state-of-the-art performance across multiple datasets.

Background & Motivation

Background: Remote photoplethysmography (rPPG) extracts physiological signals such as heart rate by capturing subtle changes in subcutaneous blood volume via ordinary cameras, and is a key technology for contactless physiological monitoring. Deep learning methods (PhysNet, PhysFormer, RhythmMamba, etc.) have become the dominant paradigm.

Limitations of Prior Work:

  1. Most existing deep learning models are designed heuristically—treating rPPG as a generic spatiotemporal signal processing task, with architecture choices relying on empirical trial and error.
  2. The lack of physical theoretical grounding means models may overfit dataset-specific noise patterns, resulting in poor cross-domain generalization.
  3. Artifacts from head motion and illumination changes are far stronger than genuine pulse signals, and "black-box" models cannot provide reliability guarantees.

Key Challenge: High-performance deep learning models vs. lack of physical interpretability and theoretical guarantees.

Goal: Can one design an rPPG model whose architecture directly embodies the physical laws governing the signal, starting from first principles?

Key Insight: Derive blood flow pulse dynamics from the Navier-Stokes equations and rigorously prove that TCN is the physically correct architectural choice.

Core Idea: The physical dynamics of rPPG signals are equivalent to causal convolution; therefore, TCN is not a heuristic choice but a physical inevitability.

Method

Overall Architecture

Visual encoder (3 EST Blocks, each containing a ZAS module) extracts spatiotemporal features → Adaptive Spatial Filter (ASF) generates spatial attention masks, aggregates features, and computes temporal differences → Gated Temporal Convolutional Network (GTCN) models long-range temporal dynamics → rPPG waveform output.

Key Designs

  1. Physics Derivation Chain: From Navier-Stokes to TCN

    • Starting point: The Beer-Lambert law establishes a linear relationship between pixel variation \(\Delta I(t)\) and subcutaneous blood volume \(\Delta V(t)\); vascular compliance further links \(\Delta V(t)\) to local blood pressure pulsation \(z(t)\).
    • Linearizing the Navier-Stokes equations → 1D momentum + continuity equations → eliminating velocity variables yields a damped wave equation: \(\frac{\partial^2 p'}{\partial t^2} + \alpha \frac{\partial p'}{\partial t} = c^2 \frac{\partial^2 p'}{\partial x^2}\)
    • At a fixed observation point \(x_0\), this degenerates to a second-order ODE (damped harmonic oscillator): \(\ddot{z} + \alpha \dot{z} + \omega^2 z = u(t)\)
    • Semi-implicit Euler discretization → LTI state-space model → Proposition 1 proves its solution is a causal convolution \(z_t = \sum_{m=0}^{\infty} g[m] \cdot a_{t-m}\) → Proposition 2 proves that an FIR filter can approximate the IIR to arbitrary precision \(\varepsilon\) → TCN is the exact computational realization of this physical process.
    • Significance: This establishes, for the first time, a complete logical chain from hemodynamic first principles to a specific network architecture.

  2. Zero-FLOPs Axial Swapper (ZAS)

    • Applies intra-block spatial transposition to the last \(k = \lfloor pC \rfloor\) channels of the feature map (dividing \(H \times W\) into \(b \times b\) blocks and performing matrix transposition), leaving the remaining channels unchanged.
    • Key properties: self-inverse (ZAS(ZAS(\(X\))) = \(X\), guaranteeing invertibility and gradient stability); energy-preserving (\(\|\)ZAS(\(X\))\(\|_2 = \|X\|_2\), 1-Lipschitz to prevent signal amplification).
    • Design Motivation: Injects cross-region spatial interaction with zero FLOPs and zero parameters, enhancing feature mixing across distant facial regions.

  3. Adaptive Spatial Filter (ASF)

    • For each frame, a lightweight convolution generates a spatial logit map → spatial softmax normalization produces an attention mask \(M_t\) → weighted aggregation over the spatial dimensions yields a 1D feature vector \(\mathbf{z}_t\).
    • Simultaneously computes a first-order temporal difference \(\mathbf{v}_t = \mathbf{z}_t - \mathbf{z}_{t-1}\) to encode pulse "velocity."
    • Output = \([\mathbf{z}_t, \mathbf{v}_t]\) channel concatenation, retaining both spatially purified intensity information and short-term temporal variation.
    • Design Motivation: The forehead and cheeks have high SNR, while other regions are dominated by noise—making global average pooling (GAP) suboptimal.

  4. Gated Temporal Convolutional Network (GTCN)

    • Dual-path causal dilated TCN: one path with tanh activation, one path with sigmoid gating → element-wise multiplication for fusion.
    • Physical significance: Implements the causal convolution operation derived in Propositions 1 & 2, modeling long-range temporal dynamics.

Loss & Training

Negative Pearson correlation loss: \(\mathcal{L}_{\text{pred}} = -\frac{\sum_t (\hat{y}_t - \bar{\hat{y}})(y_t - \bar{y})}{\sqrt{\sum_t (\hat{y}_t - \bar{\hat{y}})^2 \sum_t (y_t - \bar{y})^2}}\), directly optimizing morphological similarity between the predicted waveform and the ground truth.

Key Experimental Results

Main Results (In-Domain Evaluation)

Method UBFC MAE↓ UBFC RMSE↓ PURE MAE↓ PURE RMSE↓ BUAA MAE↓ MMPD MAE↓ Params
PhysNet 2.95 3.67 2.10 2.60 10.89 4.80 Large
PhysFormer 0.92 2.46 1.10 1.75 8.45 11.99 Large
RhythmFormer 0.50 0.78 0.27 0.47 9.19 4.69 Medium
Contrast-Phys+ 0.21 0.80 0.48 0.98 - - Medium
Style-rPPG 0.17 0.41 0.39 0.62 - - Medium
LST-rPPG 0.16 0.57 0.32 0.62 - - Medium
PHASE-Net 0.15 0.53 0.14 0.35 5.89 4.78 0.29M

Ablation Study (Cross-Domain Generalization, Leave-One-Out)

Method Others→U MAE↓ Others→P MAE↓ Others→B MAE↓ Others→M MAE↓
PhysFormer 10.29 19.75 22.09 13.90
RhythmFormer 14.71 21.11 6.04 16.14
EfficientPhys 12.87 7.15 32.30 12.87
PHASE-Net 10.04 2.86 - -

Key Findings

  • MAE of 0.14 bpm on PURE, halving that of RhythmFormer (0.27)—physical priors as inductive bias significantly improve accuracy.
  • SOTA performance achieved with only 0.29M parameters—theoretical rigor and extreme efficiency are unified.
  • Cross-domain Others→PURE MAE of 2.86 bpm, substantially outperforming PhysFormer (19.75) and RhythmFormer (21.11)—physical priors enhance generalization.
  • On challenging datasets such as BUAA and MMPD, methods like PhysFormer exhibit negative correlation (\(R < 0\)), while PHASE-Net maintains positive correlation.

Highlights & Insights

  • First derivation of an rPPG network architecture from first principles: The complete mathematical proof chain from Navier-Stokes → ODE → SSM → causal convolution → TCN elevates architectural selection from empiricism to physical inevitability.
  • ZAS zero-FLOPs module: Pure permutation operations enhance cross-region feature interaction; the mathematical proofs of self-inverse and energy-preserving properties elegantly guarantee training stability.
  • Temporal difference design in ASF: Unifies spatial aggregation and temporal differentiation in a single module, providing the downstream physical model with complete state information in the form of "position + velocity."
  • Paradigm of theoretical rigor combined with engineering minimalism: 0.29M parameters demonstrate that well-chosen inductive biases can dramatically reduce model complexity.

Limitations & Future Work

  • The physical derivation relies on several simplifying assumptions (laminar flow, linearization, single-point observation, elastic restoring force approximation), which may not hold under extreme motion or atypical vascular conditions.
  • The block size \(b\) and channel ratio \(p\) of ZAS require manual specification, lacking an adaptive mechanism.
  • Validation on large-scale in-the-wild datasets such as VIPL-HR has not been performed.
  • The cross-domain generalization table has missing entries (Others→B, Others→M), leaving the evaluation incomplete.
  • vs. PhysFormer/RhythmMamba: These methods model temporal sequences with Transformers/SSMs, representing general-purpose sequence models; PHASE-Net demonstrates from a physical perspective that causal convolution (TCN) is the correct computational primitive for rPPG.
  • vs. conventional PINN paradigm: Classical PINNs embed physical equations into the loss function; PHASE-Net's innovation lies in using physical laws to constrain the network architecture itself—"physics determines structure" rather than "physics constrains training."
  • Implication: This methodology of "deriving network structure from PDEs" is generalizable to other signal processing tasks with well-defined physical models (e.g., seismic waves, acoustic signals).

Rating

  • Novelty: ⭐⭐⭐⭐⭐ First derivation of an rPPG network architecture from first principles; paradigm-level significance.
  • Experimental Thoroughness: ⭐⭐⭐⭐ In-domain and cross-domain evaluation on 4 datasets with comprehensive ablations, though some cross-domain entries are missing.
  • Writing Quality: ⭐⭐⭐⭐⭐ Rigorous derivations and a clear, coherent logical chain from physics to architecture.
  • Value: ⭐⭐⭐⭐⭐ The physics-driven architecture design paradigm has broad applicability; the extreme efficiency of 0.29M parameters is well-suited for deployment.