OpenNovelty.org

PhysTTT: Accurate and Lightweight Cross-Domain Heart Rate Measurement with Test-Time Training

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
rPPGHeart rate measurementTest-Time TrainingCross-domain
Abstract:

Remote photoplethysmography (rPPG), a contactless technology for measuring physiological signals, holds significant promise for smart healthcare and affective computing. However, a key challenge for existing deep learning methods is the paradox between maintaining high measurement accuracy and ensuring low computational cost, especially in cross-domain scenarios. To address this, we propose PhysTTT, a novel and lightweight framework for heart rate measurement that integrates multiple 1D-CNNs with residual structures and a Test-Time Training (TTT) layer. Multi-time frame differences fusion and 1D-CNNs extract spatio-temporal features from facial video sequences by modeling subtle brightness variations, the TTT layer compresses the context information into a learnable vector space, enhancing the temporal modeling capability. Crucially, the TTT mechanism enables the model to adapt to unseen data distributions during inference, significantly boosting cross-domain generalization. Extensive experiments demonstrate that PhysTTT achieves state-of-the-art accuracy in both in-domain and cross-domain evaluations, offering an optimal balance of high performance, strong generalization, and low computational cost. Our code is publicly available at https://anonymous.4open.science/r/PhysTTT-B605/.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

PhysTTT proposes a lightweight framework combining 1D-CNNs with residual structures and a Test-Time Training layer for cross-domain heart rate measurement. The paper resides in the Domain Adaptation Methods leaf under Cross-Domain Generalization and Adaptation, alongside two sibling papers (SFDA-rPPG and Wearable Domain Adaptation). This leaf represents a focused research direction within a broader taxonomy of 50 papers across approximately 36 topics, indicating that domain adaptation for rPPG is an active but not overcrowded area. The taxonomy shows that neighboring leaves explore domain generalization without target data and meta-learning approaches, suggesting PhysTTT occupies a distinct niche emphasizing test-time personalization.

The taxonomy structure reveals that PhysTTT sits at the intersection of domain adaptation and architectural innovation. Neighboring branches include Domain Generalization Techniques (which avoid target data entirely) and Meta-Learning and Instance-Based Transfer (which enable rapid subject-specific tuning). The scope note for Domain Adaptation Methods explicitly includes test-time training, positioning PhysTTT within established boundaries. However, the paper's use of 1D-CNNs also connects to the Convolutional Neural Network-Based Methods leaf under Deep Learning Architectures, suggesting cross-cutting contributions. The taxonomy's exclude notes clarify that PhysTTT's adaptation mechanism distinguishes it from generalization-only approaches in sibling leaves.

Among 23 candidates examined, the analysis identified 4 refutable pairs across 3 contributions. The claim of first applying Test-Time Training to rPPG was examined against 10 candidates, with 3 appearing to refute this novelty assertion. The multi-dimensional feature alignment contribution faced 10 candidates, with 1 providing overlapping prior work. The core PhysTTT framework integration examined 3 candidates with no clear refutations. These statistics suggest that while the architectural combination may be novel, the individual components (TTT paradigm, feature alignment) have precedents within the limited search scope. The analysis does not claim exhaustive coverage of all rPPG literature.

Based on the limited search of 23 semantically similar papers, PhysTTT appears to offer a distinctive combination of lightweight architecture and test-time adaptation, though individual technical elements show overlap with prior work. The taxonomy context indicates the paper addresses a recognized challenge (cross-domain adaptation) in a moderately populated research area. The analysis cannot rule out additional precedents beyond the top-K semantic matches examined, and the contribution-level statistics reflect only the candidates retrieved, not the entire field.

Taxonomy

Core-task Taxonomy Papers
50
3
Claimed Contributions
23
Contribution Candidate Papers Compared
4
Refutable Paper

Research Landscape Overview

Core task: cross-domain heart rate measurement using remote photoplethysmography. The field of remote photoplethysmography (rPPG) has evolved into a rich landscape organized around several complementary research directions. Signal Extraction and Processing Methods focus on classical and hybrid techniques for isolating pulse signals from video, while Deep Learning Architectures for rPPG explore neural network designs that learn end-to-end mappings from facial video to physiological waveforms. Cross-Domain Generalization and Adaptation addresses the challenge of deploying models trained on one dataset or recording condition to new environments, a critical bottleneck given the sensitivity of rPPG to lighting, skin tone, and motion. Robustness to Environmental and Motion Challenges tackles practical issues such as illumination variation and subject movement, and Application-Specific Implementations demonstrate rPPG in contexts ranging from automotive monitoring to veterinary care. Finally, Evaluation, Benchmarking, and Tooling provides standardized datasets and software frameworks that enable reproducible comparisons across methods. Within Cross-Domain Generalization and Adaptation, a particularly active line of work investigates domain adaptation strategies that reduce the need for labeled target data. PhysTTT[0] exemplifies test-time training approaches that adapt models on-the-fly using unlabeled test videos, contrasting with methods like SFDA-rPPG[31] which perform source-free domain adaptation by leveraging pseudo-labels or self-supervision during deployment. Wearable Domain Adaptation[5] explores transferring knowledge from contact-based sensors to camera-based rPPG, highlighting the diversity of domain shift scenarios. These works collectively address the trade-off between adaptation flexibility and computational overhead, with PhysTTT[0] emphasizing real-time personalization without requiring extensive offline retraining, while neighboring approaches such as Cross-Dataset Generalization[7] focus on learning invariant representations during the training phase. Open questions remain around balancing generalization across diverse populations and the robustness of adaptation under extreme environmental conditions.

Claimed Contributions

PhysTTT framework integrating 1D-CNNs with Test-Time Training layer

The authors introduce PhysTTT, a framework that combines multiple 1D-CNNs with residual structures and a TTT layer for heart rate measurement. The TTT mechanism enables the model to adapt to unseen data distributions during inference, significantly improving cross-domain generalization while maintaining low computational cost.

3 retrieved papers
First application of Test-Time Training paradigm to rPPG research

The authors claim to be the first to apply the Test-Time Training paradigm to remote photoplethysmography (rPPG) research. This approach addresses the challenge of adapting to real data distributions in different future application domains that cannot be predicted during training time.

10 retrieved papers
Can Refute
Multi-dimensional feature alignment for fine-grained BVP waveform recovery

The authors develop a multi-dimensional alignment strategy that aligns predicted and ground truth signals across trend patterns, frequency domain, and waveform features. This approach enables more realistic fitting of BVP signal waveform characteristics for accurate physiological signal extraction.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

PhysTTT framework integrating 1D-CNNs with Test-Time Training layer

The authors introduce PhysTTT, a framework that combines multiple 1D-CNNs with residual structures and a TTT layer for heart rate measurement. The TTT mechanism enables the model to adapt to unseen data distributions during inference, significantly improving cross-domain generalization while maintaining low computational cost.

Contribution

First application of Test-Time Training paradigm to rPPG research

The authors claim to be the first to apply the Test-Time Training paradigm to remote photoplethysmography (rPPG) research. This approach addresses the challenge of adapting to real data distributions in different future application domains that cannot be predicted during training time.

Contribution

Multi-dimensional feature alignment for fine-grained BVP waveform recovery

The authors develop a multi-dimensional alignment strategy that aligns predicted and ground truth signals across trend patterns, frequency domain, and waveform features. This approach enables more realistic fitting of BVP signal waveform characteristics for accurate physiological signal extraction.

Table of Contents