Reading Notes

Personal reading notes. Original paper: arXiv:2605.23174v1.

arxiv_id: 2605.23174v1

LQ-rPPG Deep Dive: Label Quantization and Coarse-to-Fine Learning Framework

Paper Basic Information

• Paper Title: LQ-rPPG: A Label-Quantized Coarse-to-Fine Learning Framework for Remote Physiological Measurement
• Paper ID: 2605.23174v1
• Authors: Jun Seong Lee, Samyeul Noh, Changki Sung, Hyun Myung
• Affiliation: Electronics and Telecommunications Research Institute (ETRI), Korea Advanced Institute of Science and Technology (KAIST)
• Core Fields: Computer Vision, Remote Physiological Measurement (rPPG), Deep Learning, Label Learning

Research Background and Problem Definition

Remote Photoplethysmography (rPPG) technology enables non-contact measurement of physiological signals (such as Heart Rate HR and Heart Rate Variability HRV) by analyzing subtle skin color changes in facial videos caused by cardiac activity. This technology shows significant application potential in telemedicine and daily health monitoring.

In recent years, deep learning-based rPPG methods have made significant progress. From early 2D CNNs capturing spatial features, to combining RNNs and 3D CNNs to capture spatiotemporal dynamics, and further to introducing Transformer and Mamba architectures to model global context, the evolution of model architectures has significantly improved estimation accuracy and robustness.

However, existing research focuses primarily on model architecture design, often overlooking the impact of training label quality on model learning. In practical applications, contact-based PPG signals used as “ground truth” for training rPPG models are not perfect; they typically contain various forms of noise and variability:

1. Motion Artifacts: Caused by subject movement or sensor interference.
2. Inconsistent Sensor Contact: Unstable contact status between the sensor and the skin.
3. Morphological Distortions: Differences in waveform shape across different individuals or at different times for the same subject.
4. Amplitude Fluctuations: Unstable signal intensity.

This label inconsistency can lead models to overfit to noise and variability specific to the labels rather than learning core physiological patterns, thereby compromising generalization performance.

Figure 1: Examples of noise and variability in contact-based PPG signals (embedded figure from page 13 of the PDF). From left to right, it shows noise, intra-subject variability (morphology and amplitude), and inter-subject variability.

As shown in Figure 1, even for the same subject under the same settings, PPG signals exhibit differences in waveform morphology, amplitude, and noise levels. This inconsistency in the supervisory signal makes it difficult for the model to learn stable and generalizable physiological representations.

Core Method: LQ-rPPG Framework

To address the aforementioned label noise and variability issues, the authors propose LQ-rPPG, a Label-Quantized Coarse-to-Fine learning framework. This framework aims to mitigate overfitting through structured discrete labels while preserving high-fidelity waveforms.

LQ-rPPG consists of two core components:

1. Label Quantization Module (LQ)

This module converts continuous PPG signals into Multi-bit Quantized Pseudo Labels.

• Noise Reduction Mechanism: The quantization process itself has a smoothing effect, reducing high-frequency noise and minor fluctuations in continuous signals.
• Hierarchical Representation: Quantized labels with different bit depths possess different stability and fidelity. Low-bit labels are more stable but contain fewer details, while high-bit labels are rich in details but more sensitive to noise.

2. Coarse-to-Fine Estimation Model (C2F)

Based on the quantized pseudo-labels, the model adopts a hierarchical supervision strategy for training:

• Coarse-Grained Stage: First, supervision is performed using low-bit resolution (e.g., 1-bit or 2-bit) pseudo-labels, allowing the model to capture global physiological rhythms (such as heart rate frequency) and establish a stable foundational representation.
• Fine-Grained Stage: Subsequently, supervision is performed using high-bit resolution pseudo-labels to gradually refine waveform features and restore fine physiological details.

This design cleverly balances the trade-off between “stability” and “waveform fidelity”. Single-resolution discretization is either too coarse and loses details or too fine and introduces noise, whereas LQ-rPPG addresses both through hierarchical learning.

Figure 2: Schematic diagram of the LQ-rPPG framework (embedded figure from page 20 of the PDF). It shows the overall process from input video to label quantization, and then to coarse-to-fine estimation.

Experimental Verification and Performance Analysis

The authors evaluated LQ-rPPG on multiple benchmark datasets to verify its robustness and generalization capabilities under different conditions.

1. Evaluation Metrics and Datasets

The experiments covered both Intra-dataset and Cross-dataset evaluations to comprehensively test the model’s generalization performance. Key metrics include heart rate estimation error, waveform similarity, etc.

2. Performance Results

Experimental results indicate that LQ-rPPG achieves robust and generalizable rPPG estimation under various challenging conditions. Through hierarchical supervision, the model effectively reduces overfitting to label-specific variations and learns structured and consistent representations.

Figure 3: Comparative analysis of experimental results (embedded figure from page 30 of the PDF). Likely shows performance comparison of different methods on major datasets.

Figure 4: Comparative analysis of experimental results (embedded figure from page 30 of the PDF). Likely shows waveform estimation effects in specific scenarios.

3. Efficiency Advantages

In addition to accuracy improvements, LQ-rPPG performs excellently in computational efficiency, which is crucial for practical deployment:

• Parameter Reduction: Compared to existing methods, the number of parameters is reduced by 88%.
• Computational Load Reduction: Multiply-Accumulate Operations (MACs) are reduced by 29%.
• Throughput Increase: Processing throughput is increased by 191%.

This indicates that LQ-rPPG is not only more accurate but also lighter and more efficient, making it suitable for remote health monitoring on resource-constrained devices.

Figure 5: Comparative analysis of experimental results (embedded figure from page 30 of the PDF). Likely shows efficiency metrics or ablation study results.

Figure 6: Comparative analysis of experimental results (embedded figure from page 30 of the PDF). Likely shows specific data on cross-dataset generalization capabilities.

Summary and Outlook

LQ-rPPG proposes a novel perspective, addressing the robustness problem in rPPG estimation by starting from label quality. By introducing label quantization and a coarse-to-fine learning strategy, this framework successfully mitigates the negative impact of noise and variability in contact-based PPG labels on model training.

Summary of Key Contributions:

1. Problem Insight: Identified the critical issue that existing rPPG methods overlook the quality of training labels and its impact on generalization performance.
2. Methodological Innovation: Proposed the LQ-rPPG framework, combining a label quantization module and a hierarchical supervision model to achieve a balance between stability and high fidelity.
3. Performance Verification: Demonstrated the superiority of the method on multiple benchmark datasets, while achieving significant improvements in model efficiency and throughput.

This research provides a new direction for future rPPG studies, indicating that besides optimizing network architectures, optimizing the structure and quality of supervisory signals is also a key path to improving model performance. The code has been made public, facilitating further reproduction and improvement by the community.