Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Sep 4·5 min read
Cross-Dataset Transfer and Reliability of Explainable Artificial Intelligence for RhythmFormer Remote Photoplethysmography
TL;DR
RhythmFormer models were developed for remote photoplethysmography, estimating heart rates from facial videos. The study found that explanation methods varied in reliability across datasets, with Beyond Intuition performing best overall.
✦ Why It Matters
Consider using Beyond Intuition for more reliable model explanations in your rPPG applications.
Key Takeaways
- Explore different explanation methods for AI models in rPPG.
- Utilize Beyond Intuition for better model interpretability.
- Assess model performance using multiple datasets for robustness.
- Be cautious of relying solely on explanation metrics for accuracy.
Full Summary
Remote photoplethysmography (rPPG) is a technique that estimates cardiovascular pulse from facial video. This study focused on the RhythmFormer models, which were trained on two datasets: NCKU-rPPG and UBFC-rPPG, under various conditions like lighting and movement.
The researchers evaluated different explanation methods, including raw attention and Beyond Intuition, using metrics such as skin coverage and the Salience-guided Faithfulness Coefficient (SaCo). Results showed that Beyond Intuition had the highest median coverage (0.789) and SaCo (0.837) on the NCKU-rPPG dataset, while also performing well on UBFC-rPPG.
However, the correlation between explanation metrics and heart rate accuracy was weak, indicating that the explanations did not reliably track model performance across conditions. These findings suggest that while some explanation methods are more effective, their utility in assessing model performance may be limited.