Whether to recalibrate depends on whether drift exceeds the acceptable error

What it is

A mapping does not stay the one from calibration time. Whether to interrupt the user for a new calibration session depends on whether current error has left this UI’s acceptable range—usually set from the smallest target’s visual angle, hit spacing, and task consequence. Inside the budget, do not pull people to look at points for a prettier calibration; once the smallest targets miss routinely, even a pretty old mapping should be voided. This is a criterion on the calibration workflow, not an account of how drift is produced, and not a tutorial on implicit correction.

Why it happens

The acceptable range comes from geometry, not from the instrument brochure’s nominal accuracy. If the smallest button is 1.5° wide with 0.5° gaps, a 1° mean offset already hits neighbors systematically. If buttons are 4° and only toggle music, the same 1° can keep going. Consequence tightens the same geometry: a payment button, even if large, should not keep an old calibration once error has biased onto a neighbor.

The criterion needs an observable proxy because true gaze is usually unavailable. Proxies include: the user missing in the same direction repeatedly, residual when pointing at a known control, an on-the-spot remeasure on validation points. Treating “force calibration every N minutes” as the criterion interrupts while error is still small and is too late in the minute when error has already blown.

Studying it

Set several error budgets on the same UI (for example 0.5°, 1°, 2°), simulate over-budget with an injected offset, and watch hit rate, the judgment “I should go calibrate,” and time lost to the interruption. Work such as Hornof and Halverson’s shows that systematic error can be measured and used to decide whether to fix. Count “user requested calibration” and “system decided it was time” separately: many people still will not go look at points when misses are already obvious. Tolerance for interrupted calibration differs between assistive users and office users; budget and interruption cost should be stratified.

Where it stops holding

With no minimum target, only presence detection (looking at the screen or not), almost any old calibration is in budget and should not interrupt. Clinical or research records that must write accuracy into a paper have a methodological budget, not a button-size budget, often far stricter than the product. Calibration-free schemes have no “do another calibration session” to take; the criterion becomes “when to abandon gaze and use head direction.” Children are worn down by frequent calibration; the budget should be wider or the session gamified, but not so wide that assistive controls they depend on cannot be hit.

Applying it

• Write an error budget from smallest-target visual angle and consequence; enter a calibration session only over budget, and do not nag on a timer under it.
• Offer a lightweight on-the-spot remeasure (look at one or two known points) as a proxy; escalate to the full lattice only if that fails.
• Verify by injecting systematic offsets of different sizes and checking that the system interrupts only when hit rate becomes task-unacceptable, and that users understand why they must look at points again.

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 · C8.08.4 Calibration-free eye tracking is usually less accurate than per-user calibration
• Adjacent: C8.10 Gaze drift and recalibration · C8.09 Precision versus accuracy
• Search: error budget · recalibration criterion · gaze accuracy