Eye Scanner

Runs the trained eye-fatigue models right inside your browser. Your camera images never leave this device.

Loading models…

Upload photo

Box size

Get close so one eye fills the box, like the photos the model learned from. Good light helps.

Eye open–

Stage 1 CNN · 97.7% accurate on people it never saw

Fatigue–

–

Alert / strained / fatigued

3-way model · 51.5% (vs 38.8% guessing) on days it never saw

What the model sees: centre square, grayscale, 64×64. If this doesn't look like an eye, move closer.

Fatigue uses DINOv2 (a large pretrained vision model) on the colour image, with a classifier trained on 120 labelled phone photos plus 2,632 reviewed web photos. Alert vs fatigued scored 72.5% (vs 53% guessing) on days it never saw. It's a weak signal, not a diagnosis. One scan takes a few seconds in the browser; live mode averages the last 10 scans.

Blink & PERCLOS meter

Checks your eye about 15–20 times a second with the open/closed model (97.7%) and measures how your eyes close over the last 60 seconds. PERCLOS, the share of time the eye is closed, is the best-proven eye signal of drowsiness in sleep research.

PERCLOS–

Blinks per minute–

Average blink length–

Long closures (≥ 0.5 s)–

Higher PERCLOS, longer blinks and more long closures all point toward drowsiness. Everyone blinks differently, so compare against your own rested baseline: run a 60-second test when you're well rested and save it. Hold still, keep one eye in the box, and look at the screen. The model calls the eye closed below 30% open and open again above 70%.

Teach the model

Scan your eye, say how you actually feel, then see what the model guessed. Your answer is saved in this browser only (no photo, just the model's numbers for it), and the model learns from it when you retrain. You answer first, so the guess can't influence you.

Tips: label at different times (right after waking, late evening, after screen time) in similar light. A score worth trusting needs at least 20–30 labels of each kind over several days.