CleanerPPG dataset
2020 · 0.7 GB
Hand-verified heart beat annotations for the ground truth of 11 public rPPG datasets.
Request accessCleanerPPG is a meta-dataset with cleaned ground truth for existing rPPG datasets. Every peak (heart beat) in the ECG and PPG ground truth of the datasets listed below was verified and corrected by hand by expert annotators, for a more accurate and standardised evaluation.
Candidate peaks were first found with a gradient-based peak detector. An expert then matched, verified and corrected every candidate peak by looking at the shape of the raw PPG or ECG waveform and at the inter-beat intervals, using a purpose-built Python tool. Peaks were annotated at the crest of the PPG signal, or at the highest point of the R peak in the ECG. Ectopic beats in the ECG were included for completeness. Parts too noisy to make out clear peaks, for example from a moving electrode or finger, were left blank.
This removes false positive and false negative peaks caused by artefacts, giving noise-free heart beat annotations for an extensive set of public rPPG datasets. Annotators spent about 30 seconds per minute of signal: 36 person-hours to clean 75 hours of ground truth. The annotations allow a more accurate evaluation of rPPG methods, especially for noise-sensitive measures like HRV, and can be used to train supervised methods that tell true heart beat peaks from noise.
CleanerPPG contains annotations only. To use it, you also need the original datasets, which are available from their own authors.
In the zip: one folder per dataset, with Raw (automatically detected peaks) and Cleaned (hand-annotated peaks), or one such pair per device (e.g. ECG and PPG). Each file is a CSV with the columns Time (ms, starting at 0 at the start of the video), Signal (the original ground truth signal) and Peaks. File names follow the original dataset’s folder structure, with the folders joined by _ (or - where the original names already contain underscores). Where one ground truth file belonged to several videos, it was copied once per video. Notes per dataset:
- MMSE-HR: the original has no ECG waveform, only the instantaneous heart rate, so these files have the columns Time and HeartRate.
- MAHNOB-HCI: 30 seconds (frames 306–2135) of 527 videos, as described in Gudi et al. (arXiv).
- DEAP: only subjects with videos (s01–s22).
- UBFC-RPPG: both parts (Dataset_1 with 10 videos, Dataset_2 with 42); only Dataset_2 was used in the paper.
- VIPL-HR: sources 1–3; source 4 is not relevant for RGB video and was not cleaned.
- MoLi-PPG: timestamps interpolated between the first and last original timestamp, as the original files have several values per timestamp.
The zip includes a README with all details.
At a glance
| Type | Meta-dataset: cleaned ground truth annotations, no video |
|---|---|
| Covers | PURE, UBFC-RPPG, MMSE-HR, VIPL-HR, ECG-Fitness, MAHNOB-HCI, VicarPPG, VicarPPG-2, DEAP, MoLi-PPG, COHFACE |
| Annotations | Every peak in the ECG and PPG ground truth, verified and corrected by hand |
| Ground truth cleaned | 75 hours (36 person-hours of annotation) |
How to cite
Please cite this paper in any work that uses CleanerPPG:
Gudi A, Bittner M, van Gemert J. Real-Time Webcam Heart-Rate and Variability Estimation with Clean Ground Truth for Evaluation. Applied Sciences. 2020; 10(23):8630. MDPI · arXiv
BibTeX
@article{gudi2020real,
title={Real-Time Webcam Heart-Rate and Variability Estimation with Clean Ground Truth for Evaluation},
author={Gudi, Amogh and Bittner, Marian and van Gemert, Jan},
journal={Applied Sciences},
volume={10},
number={23},
pages={8630},
year={2020},
publisher={Multidisciplinary Digital Publishing Institute}
}