How CanthalTilt.net Measures Canthal Tilt

What canthal tilt is

Canthal tilt is the inclination of the line joining the inner and outer corners of an eye. In a published photographic method, that inclination was taken against a horizontal perpendicular to the facial midline.^[1]

Inner canthus and outer canthus

The inner, or medial, canthus is the corner next to the nose. The outer, or lateral, canthus is the corner toward the ear. The angle uses those two points. It does not trace the eyelid crease or the iris.

What the landmark model does

MediaPipe Face Landmarker finds a set of facial points in the browser, including both eye corners and iris centers.^[2] The model file and its WebAssembly runtime are loaded from this site, and only after you choose Upload a Photo, Use Camera, or Try an Example. A normal visit to a guide page does not download the model.

How the corners are found

The subject’s right outer and inner corners are landmark indices 33 and 133. The subject’s left outer and inner corners are 263 and 362. Those indices follow the model’s named eye contours. “Left” and “right” are the person’s left and right, not the left and right of the screen.

The face-relative horizontal

The primary reference is a midline fit through the forehead, the nose bridge, the base of the nose, and the chin. The horizontal is perpendicular to that axis. Several points are used so one expression-sensitive point does not set the line.

Head roll

Roll is the in-plane lean of the head. If MediaPipe’s facial transformation matrix reports a roll close to the midline roll, that matrix roll is used. If they disagree, the midline wins. The line between the two inner corners is not the reference: those corners can sit at different heights, and forcing them level would delete a real asymmetry.^[1] The pupil line is only a fallback when the midline cannot be fit.

Why a large pitch or yaw is rejected

Looking up, down, or sideways changes perspective. A single roll correction does not undo that. Frames with too much pitch or yaw are rejected so the tool does not print a precise number from a pose it cannot level.

Left, right, and mirrored selfies

A front camera may show a mirrored preview. The measurement still names the subject’s left eye and the subject’s right eye. Horizontal mirroring does not flip a positive angle into a negative one, because the sign comes from which corner is higher, not from which side of the screen the eye occupies.

A photo is not a permanent body measurement

The number describes that image, or the stable camera frames from that sitting. Another photo, another expression, or another camera height can move it. Camera height, pitch, yaw, roll, and expression can change the apparent angle.

Limits of automatic landmarks

Hair, glasses, squinting, blur, and unusual lighting can move a corner onto a lash or a shadow. The quality mark says when the photo is weak. It is not a claimed error of ±0.13° or any other invented tolerance.

Why you can move the points

Adjust eye corners lets you drag the four canthi. The angle is recalculated from the points you place, using the roll already estimated for that photo. Reset to automatic restores the model’s points. The share card uses the numbers on screen, including a manual correction, and it does not include the photo.

How the labels are defined

The signed angle is always shown first. Positive means the outer corner is higher. Negative means it is lower. The display band calls −2° through +2° near level so a tiny error does not flip the word. −1.5° remains a negative direction inside that band. The band is not a medical class. One surgical series used 0° and 5° boundaries on a sample of 76 Korean patients, and those boundaries are not reused here.^[1]

Privacy

The photo, the landmarks, and the angle stay in the browser for a standard measurement. They are not sent to an analytics endpoint. Details are on the privacy page.

Known limitations

• One photograph is not a 3D anthropometric exam.
• Iris and canthus landmarks can fail when the eyes are closed or covered.
• The midline can shift if the face is turned, which is why large yaw is rejected.
• Camera mode summarizes repeated frames. It does not create a clinical confidence interval.
• Nothing on this site is a diagnosis or a recommendation for surgery.