Calibration has the user look at known screen points to build an individual mapping

What it is

A video eye tracker sees pupil and corneal reflection in the camera, not gaze coordinates on a display. Calibration has the user look at several known points (often 5, 9, or 13) and estimates an individual mapping from those pairs: pupil–corneal-reflection vector to display plane. Without it the system has an eye image and not “where they are looking.” It is building a model, not measuring how far that model later drifted, and not defining the words precision and accuracy.

Why it happens

The visual axis (the direction actually aimed) does not coincide with the optical axis of the eyeball; the angle, kappa, differs by person. Corneal curvature, pupil-center offset in the iris, and spectacle refraction all change camera geometry. A polynomial or homography fit of observed vectors at the calibration points to known coordinates absorbs those individual constants into coefficients. Too few points, and extrapolation at the screen edge explodes; too many, and users fatigue, later fixations get worse, and the fit degrades.

Calibration quality depends on whether the user is really looking at the point, whether the head moved between points, and whether IR reflections were scrambled by lenses. A good calibration is a mapping for “now, this pose, this lighting,” not a permanent physiological constant.

Studying it

Treat number of points, their spatial layout, whether small head motion is allowed, and whether low-quality points are retried as independent variables, and immediately after calibration measure angular error on a held-out validation set. Holmqvist and colleagues’ methodology treats post-calibration validation as a precondition for reporting accuracy. Also log failure rate: how many people still miss the error threshold after N attempts, and whether failures cluster on glasses, colored contacts, or eyelid morphology. A lab of bare-eyed students writes calibration as “everyone passes on the first try.”

Where it stops holding

Some consumer devices claim to skip calibration; that is a different tradeoff, usually a population model plus in-run adaptation, with larger individual error. Strabismus, a prosthetic eye, or severe ptosis breaks “look at the points in order” and needs monocular calibration or giving up on gaze. Children struggle to hold a point long enough; gamified calibration raises completion, not necessarily fixation quality on the points. In bright light the pupil shrinks hard, or sun drowns the IR, and more points will not yield a stable mapping.

Applying it

• Calibrate with a lattice that covers the screen region actually used; if targets appear at the edge, put points at the edge. Retry a bad point rather than voiding the whole session.
• During calibration, detect “are they really looking at this point” (stable fixation, visible pupil) and do not write an unstable point into the fit.
• Verify by reporting angular error on held-out points immediately after calibration, stratified by glasses/bare eye and indoor/window, not only “calibration complete.”

Related

• Same group: C8.08.2 Calibration accuracy decays when head pose, glasses, or lighting change after calibration · C8.08.3 Whether to recalibrate depends on whether drift exceeds the acceptable error · C8.08.4 Calibration-free eye tracking is usually less accurate than per-user calibration
• Adjacent: C8.09 Precision versus accuracy · C8.10 Gaze drift and recalibration
• Search: calibration · pupil-CR · kappa angle