Implicit recalibration can correct drift from interaction with known targets without interrupting the task

What it is

Implicit recalibration does not pull the user to calibration points. When they click, pinch, or hit a control whose position is known, it treats the gaze estimate at that instant and the control’s coordinate as a new pair, infers the current offset, and patches the mapping. The task does not stop; the points are hidden inside real actions. Hornof and Halverson’s use of click residuals to clean systematic error is an early form. It is an online patch after drift has started, not the first mapping, and not the switch to a full calibration once error is too large to patch.

Why it happens

A point-plus-confirm interaction yields a weak label: people are usually looking at the object they mean to hit. Estimate minus object center is a residual vector. Treat the residual as an observation of offset and update the mapping’s translation (sometimes a mild scale) with a low-pass or a short-window regression. Because the label is weak—they may be looking at the caption, not the button center, or they may have clicked the wrong thing—the update must be conservative; one mis-click must not drag the whole mapping away.

The gain is that compensation follows drift and nobody feels “go calibrate.” The cost: viewing stretches with no interaction (a film, reading) yield no labels; if interaction clusters on one region of the screen, the patch is valid there and the corners stay biased.

Studying it

In a session with injected slow drift, contrast no correction, implicit click correction, and periodic forced calibration. Report task hits, whether users noticed calibration, and how often the mapping was polluted by a mis-click. Label quality is the key independent variable: update only on high-confidence confirms versus on every contact. Test the “looking at the control center” assumption: put some clicks on edge text of a large button and see whether the patch is dragged by text fixations. A lab that only offers a small cross to click will overestimate label purity.

Where it stops holding

Users with no second modality, only dwell, lack a clean timestamp for implicit correction; the dwell itself may already be a false selection, and using it as a label reinforces the error. Rapid game clicks and dense accidental touches pollute faster than they correct. The channel should be off when children mash. Viewing apps that almost never click have no samples; implicit correction cannot replace a mid-session remeasure. On privacy, using clicks as labels pairs interaction coordinates with gaze and should be retained as sensitive data.

Applying it

• In a product that already does gaze-point plus confirm, use residuals from high-confidence confirms as a slow translation patch, with a step cap so one mis-click cannot ruin the mapping.
• Mark the region the patch is valid for; if clicks are all central, do not claim the corners were implicitly calibrated.
• Verify by injecting slow drift and confirming hit rate recovers without interruption, and by injecting mis-clicks and confirming the mapping does not run away.

Related

• Same group: C8.10.1 Drift is the gradual offset of gaze estimates from true position after calibration · C8.10.2 Head motion, glasses slippage, and posture change all worsen drift · C8.10.4 Beyond a point, implicit correction is not enough and the user must actively recalibrate
• Adjacent: C8.04 Gaze-plus-confirm combinations · C8.08 Eye-tracking calibration
• Search: implicit calibration · click correction · online gaze adaptation