Calibration-free eye tracking is usually less accurate than per-user calibration

What it is

Calibration-free schemes do not ask the user to look at a set of known points. They emit a gaze estimate from an appearance model trained on a population, an anatomically average eyeball, or a weak in-run constraint such as “people are probably looking at the screen.” Skipping the session buys this: individual kappa, corneal curvature, and spectacle refraction are not absorbed into private coefficients, so spatial error is usually larger than per-person calibration on the same hardware. That is enough for “are they looking at this region,” not for pinning a cursor on a small button.

Why it happens

An appearance model regresses over many faces and gaze directions and learns a population conditional expectation. People inside the training distribution (common skin tone, eye shape, lighting) sit near that expectation; uncommon frames, marked esotropia, makeup hiding the lid fold, sit off it. Anatomical models use mean eyeball radius and mean kappa and are biased for the individual. In-run adaptation (clicks as weak labels) can pull some of the bias back, but that is already implicit calibration without asking the user to look at points.

Per-person calibration trades a dozen seconds for coefficients that belong to this person; calibration-free spreads those seconds into larger error across the whole session. “Works out of the box” in marketing is a claim about flow, not about angular error.

Studying it

On the same camera and the same people, contrast error after 9-point calibration, fully calibration-free error, and error after several click-adaptation steps. Stratify glasses, eyelid morphology, and lighting, because calibration-free failure is not uniform. Appearance papers often report mean angular error; a product needs the tail: the worst 10% of users decide whether small targets work. Also report abandonment of the calibration session—sometimes the advantage of calibration-free is not mean error but “nobody gets stuck on the calibration screen.”

Where it stops holding

Large targets, presence detection, and a large foveation island often absorb the extra error, and calibration-free is enough. Assistive gaze typing, pixel-level annotation, and small-icon clicks still prefer per-person calibration. Children and some users with cognitive disabilities cannot finish a look-at-points task; calibration-free may be the only path that boots, even at larger error. If the model collapses on lighting or eye shapes unseen in deployment, calibration-free can be worse than a poor calibration. Some “calibration-free” headsets still had camera extrinsics calibrated at the factory; they skip the user’s points, not all calibration.

Applying it

• Choose by target size: region-level can be calibration-free; control-level should offer a short calibration and remember those coefficients.
• Do not write out-of-the-box as “as accurate as calibrated”; show the current estimated error magnitude in settings.
• Verify on people held out from training (including glasses and varied eye shapes) by reporting both median error and the worst 10%, and confirm calibration-free is only used on targets the tail can still hit.

Related

• Same group: C8.08.1 Calibration has the user look at known screen points to build an individual mapping · 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
• Adjacent: C8.09 Precision versus accuracy · C8.10 Gaze drift and recalibration
• Search: calibration-free · appearance-based gaze · person-independent gaze