IrisVision

Computer Vision · Biometrics · Runs 100% in-browser

Find the iris.
Then prove it.

IrisVision localizes the human iris in real time from a webcam or photo, fits a sub-pixel circle to it, and then independently validates that fit with classical computer vision — Daugman's integro-differential operator, Hough circles and radial intensity contrast — before reporting a verdict.

Launch live demo How it works
  • 4783D face landmarks
  • ~30frames / second
  • 3independent validators
  • 0bytes uploaded

Live demo

Iris localization lab

Start your camera or drop a face photo. Processing happens on your device — no frame ever leaves the browser.

Waiting for input

Start the camera, try the sample photo, or drag & drop a face photo here.

Loading models…

Landmarker OpenCV.js No face
— fps
Drop image to analyze

Right eye (subject)

—
Center
—
Radius
—
Fit RMS
—
Contrast
—
Daugman Δ
—
Hough Δ
—

Left eye (subject)

—
Center
—
Radius
—
Fit RMS
—
Contrast
—
Daugman Δ
—
Hough Δ
—

Pipeline

From pixels to a validated iris

Every frame flows through six stages. A learned model proposes; classical vision verifies.

  1. 01

    Capture

    Frames come from getUserMedia or a local image. Nothing is uploaded — all compute runs in WebAssembly and WebGL.

  2. 02

    Face landmarks

    MediaPipe Face Landmarker regresses 478 3D landmarks, including 5 points per iris (center + 4 limbus points).

  3. 03

    Circle fit

    A centered algebraic least-squares (Kåsa) fit over the limbus points gives a sub-pixel center, radius and RMS residual.

    min Σ (x² + y² + Dx + Ey + F)²
  4. 04

    Eye state

    Eye Aspect Ratio detects blinks with hysteresis; the iris position between canthi and lids yields a gaze estimate.

    EAR = (‖p₂−p₆‖ + ‖p₃−p₅‖) / 2‖p₁−p₄‖
  5. 05

    Validation

    The eye region is cropped and re-examined with Daugman's integro-differential operator, OpenCV Hough circles and a radial intensity contrast test.

  6. 06

    Metrics

    Using the near-constant human iris diameter (~11.7 mm) and a pinhole model, the system estimates camera distance.

    Z = f · 11.7 mm / dpx

Validation

Why a second opinion matters

Landmark models always output an answer — even when the eye is closed, blurred or occluded. IrisVision only trusts a fit that agrees with evidence in the pixels.

Geometric agreement 45%

Two classical detectors re-locate the limbus from raw pixels: Daugman's integro-differential operator on the lateral arcs the eyelids rarely cover (65%), and OpenCV HoughCircles on an upsampled, median-filtered window (35%). Each is scored on center error and radius ratio.

max(x,y,r) | ∂/∂r ∮ I(x,y) ds / 2πr | g = (1 − Δc/0.6r) · (1 − |1 − r′/r|/0.5)

Radial contrast 35%

The iris is darker than the surrounding sclera. Mean intensity inside 0.7r is compared to a 1.25r–1.9r ring, sampled only inside the eyelid polygon.

k = (Iring − Iiris) / Iring

Eye openness 20%

A closed or squinting eye hides the limbus. Low EAR marks the eye as OCCLUDED rather than pretending to localize it.

o = clamp((EAR − 0.15) / 0.12)
PASS score ≥ 0.60 FAIL score < 0.60, with the weakest signal as reason OCCLUDED EAR < 0.17

Built with