Data
A features-only dataset
Per-frame eye features and hand-made labels of spontaneous and deliberate eye closures, recorded with consumer webcams, so that anyone can test their own blink detectors. No images, no video, no faces.
Not released yet. The dataset will be published in
data/public/ on GitHub
once Study A has been labelled, together with a full datasheet
(data/DATASHEET.md).
What is in it
Per-frame features (one row per video frame)
| Column | Meaning |
|---|---|
t_ms | Frame capture time in milliseconds |
face | Whether a face was found in the frame |
earL, earR | Eye aspect ratio of each eye, from six landmarks |
blinkL, blinkR | MediaPipe blink scores (blendshapes), 0 to 1 |
yaw, pitch, roll | Head pose in degrees |
n_faces | Number of faces in view |
eye_pts | The 12 eye landmarks (6 per eye) as normalised x, y; no other face points |
Labels (one row per closure)
Made by hand by the experimenter in the labelling tool. Only closures confirmed by a person count as ground truth.
| Column | Meaning |
|---|---|
onset_ms, offset_ms | Start and end of the closure, on the frame clock |
type | spontaneous, deliberate-short, deliberate-long, wink-L or wink-R |
labeller, labelled_at | Who labelled it, and when |
Session information
Cue times and types, block start and end times, the participant ID (P01, P02, …), age range, glasses (yes or no), lighting, camera model and the measured frame rate.
What is not in it
- No images or video, and no face mesh beyond the 12 eye points.
- No names, contact details or anything else that identifies a participant.
- No audio.
Datasheet in brief
- Purpose
- Measure how well a consumer webcam separates deliberate from spontaneous blinks, and how much per-user calibration helps.
- Who
- Adult volunteers (18+), able-bodied, mostly a convenience sample. The planned size is at least 8 people.
- How
- Laptop webcams at about 50–70 cm in normal room light; about 12 minutes per person in the app’s study mode, with everything recorded on the experimenter’s laptop.
- Labelling
- Every closure labelled by hand by the experimenter; two participants re-labelled a week later to measure self-agreement.
- Intended uses
- Research on blink detection, calibration and blink-based text entry.
- Not intended for
- Identifying people, or any clinical or diagnostic use.
- Licence
- Creative Commons Attribution 4.0 (CC BY 4.0). The code is MIT-licensed.
Privacy and consent
- Every participant gives consent on a screen in the app before anything is recorded.
- Video is recorded only if the participant ticks a separate box, only to help with labelling, and it is deleted after labelling unless the participant agreed otherwise. Video is never published.
- Participants can withdraw at any time and have their data deleted.
- Files carry an ID only. Any link between an ID and a name is kept offline, on paper.
- Participation is unpaid.
- The public website and the Try it app record nothing: frames are processed in your browser and never uploaded.