Camera-Based Remote Physiology Sensing for Hundreds of Subjects Across Skin Tones

Jiankai Tang,Xinyi Li,Jiacheng Liu,Xiyuxing Zhang,Zeyu Wang,Yuntao Wang

2024-04-07

Abstract:Remote photoplethysmography (rPPG) emerges as a promising method for non-invasive, convenient measurement of vital signs, utilizing the widespread presence of cameras. Despite advancements, existing datasets fall short in terms of size and diversity, limiting comprehensive evaluation under diverse conditions. This paper presents an in-depth analysis of the VitalVideo dataset, the largest real-world rPPG dataset to date, encompassing 893 subjects and 6 Fitzpatrick skin tones. Our experimentation with six unsupervised methods and three supervised models demonstrates that datasets comprising a few hundred subjects(i.e., 300 for UBFC-rPPG, 500 for PURE, and 700 for MMPD-Simple) are sufficient for effective rPPG model training. Our findings highlight the importance of diversity and consistency in skin tones for precise performance evaluation across different datasets.

Computer Vision and Pattern Recognition,Artificial Intelligence

What problem does this paper attempt to address?

The paper attempts to address the following key issues:

1. **Reliability and Diversity of Datasets**: Existing remote photoplethysmography (rPPG) datasets are lacking in scale and diversity, limiting the ability to conduct comprehensive evaluations under different conditions. The paper addresses this issue by analyzing the largest real-world rPPG dataset, VitalVideo.

2. **Effectiveness of Model Training**: The study finds that datasets containing a few hundred individuals (e.g., 300 individuals in UBFC-rPPG, 500 individuals in PURE, and 700 individuals in MMPD-Simple) are sufficient to effectively train rPPG models. This indicates that small, well-curated datasets can also achieve efficient training.

3. **Impact of Skin Tone**: The paper emphasizes the importance of skin tone diversity and consistency on model performance, especially in cases of darker skin tones. Experimental results show that maintaining consistent skin tones in training and testing sets can significantly improve testing accuracy.

4. **Relationship Between Data Volume and Model Performance**: The study finds that increasing the amount of training data does not always enhance model performance. There is a threshold beyond which adding more samples has little effect on performance improvement.

5. **Cross-Dataset Validation**: The paper validates the model's generalization ability under different conditions through cross-dataset training and testing (e.g., PURE, UBFC-rPPG, and MMPD-Simple), and emphasizes the importance of dataset diversity.

Through these studies, the paper aims to advance camera-based physiological signal monitoring technology, enabling it to accurately measure vital signs in more diverse environments.