Diverse rPPG Dataset (rPPG-26)

Remote photoplethysmography (rPPG) is an emerging contactless technique for physiological monitoring. However, the robustness of signal extraction is often challenged by environmental variability, demographic differences, and dataset biases. Standard datasets such as UBFC, PURE, and COHFACE exhibit limited variability for real-world applications since they are largely collected under controlled indoor conditions. Consequently, regions with higher melanin levels, diverse facial tones, and complex lighting remain underrepresented. To address this gap, we introduce a new rPPG dataset featuring recordings across diverse skin tones, lighting environments, and realistic facial expressions in natural scenarios. We further augment the dataset by integrating MediaPipe’s 468-landmark tracking, unsupervised feature learning (FactorizePhys), and TSFEL-based feature engineering for emotion-related analysis. Using this enhanced framework, we conducted cross-dataset evaluations on the UBFC dataset employing standard rPPG-Toolbox pipelines. The inclusion of MediaPipe landmarks reduced mean absolute error (MAE) from 13.86 to 3.42 BPM (ICA) and from 18.98 to 2.63 BPM (LGI), reflecting strong cross-domain generalizability. FactorizePhys demonstrated notable stability under noise and compression, while bias-immune preprocessing yielded six-class emotion recognition accuracy of up to 63% using the ICA method. Overall, this study contributes a demographically rich and environmentally diverse dataset, alongside enhanced feature extraction and benchmarking results, to advance equitable and robust rPPG-based physiological and emotional monitoring in real-world contexts.

Verbatim fromSemantic Scholar· Abstract via Semantic Scholar; © the authors/publisher