Reproduce “MMPD: Multi-Domain Mobile Video Physiology Dataset” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiankai Tang, Kequan Chen, Yuntao Wang, Yuanchun Shi, Shwetak Patel, Daniel McDuff, Xin Liu
The paper introduces MMPD, a multi-domain mobile video physiology dataset, comprising 11 hours of recordings across diverse skin tones (Fitzpatrick 3-6), natural motions, and varied lighting, all captured with mobile phones. Evaluations reveal that current rPPG algorithms struggle significantly with darker skin tones, walking motions, and natural light, highlighting limitations in existing models and training data.
Key insights
- Both unsupervised signal processing and supervised deep learning rPPG methods show significant generalizability issues and performance degradation when tested on the diverse conditions within MMPD.
- Current rPPG algorithms struggle substantially with subjects having darker skin tones (Fitzpatrick types 4-6), yielding high error rates and poor correlation.
- Naturalistic motion, especially walking, and natural lighting conditions pose major challenges for existing rPPG techniques, demonstrating a clear gap in robust performance for real-world applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.