Why Ratios, Not Millimetres: Iris-Distance Normalisation Explained
A photo has pixels, not centimetres. Without a known scale, such as a ruler in frame or camera calibration, you cannot say how wide a face is in millimetres. What you can do is compare parts of the face to each other.
The normaliser
Selfbaus divides geometry values by the distance between the two iris centers in the image. The mesh provides an iris center point for each eye (points 468 and 473 in the MediaPipe numbering). The result is expressed in "IPD ratio" units: a face width of 3.0 means the width equals three iris-center distances in that photo.
This removes overall size, so a face photographed close and a face photographed far give comparable numbers, to a first approximation. We use the iris distance only as a divisor. We never convert it to millimetres.
What it does not fix
- Perspective: a close camera magnifies the nose relative to the ears (see the camera distance article), changing ratios.
- Pose: turning the head shrinks the apparent distance between the eyes.
- Expression and gaze: squinting and strong eye movement shift the iris points a little.
- Model error: the iris center is a prediction.
The three display-only ratios and the overlay
Besides symmetry and the vertical segments, the report shows three capture ratios. Each is a distance between mesh points divided by something measured in the same photo:
- Eye spacing: the gap between the inner eye corners (points 133 and 362) divided by the mean of the two eye widths (33 to 133 and 362 to 263).
- Mouth width: mouth corners (61 and 291) divided by the iris-center distance (468 to 473).
- Nasal contour width: points 129 and 358 divided by the iris-center distance. These are contour points on the mesh, not the alar base. The mesh has no true alar-base landmark, so this is an estimate of nose width in this frame, not measured anatomy.
They are descriptive. We show no norms, no percentile, no "good" range and they do not feed any score. Expression changes mouth and eye geometry, and the limits above (perspective, pose, model error) apply to all three.
A toggle draws the points and lines on your own photo, on your device. It uses the same EXIF-oriented pixels that were analyzed, and each line is a numerator or denominator from the list above. The photo is never uploaded and is not written to storage. Each completed check saves its ratios to your browser's local storage on this device only (see the privacy page), and the Plan page can clear them. The overlay is hidden when a photo fails the frame check.
How stable are the numbers? We ran a small check on four images (two AI-generated adults, each in two lighting conditions) with nine image transforms each. The three ratios moved by up to 5.0%, 2.4% and 4.1%, and a 1 degree rotation sometimes triggered a retake instead of ratios. In two AI-edited head-turn images inside the gate, the nose ratio moved 2-3%. That is observed spread on synthetic faces. It is not a population validation, not a repeated-capture study and it gives no calibrated uncertainty.
Why this is still useful
Because every measurement is divided by the same quantity, the numbers are internally consistent and repeatable for the same photo. They describe that frame. They should not be read as anatomical constants.
Practice
If you compare ratios across photos, keep pose, distance and expression the same, and treat small differences as noise.