Beyond a point, implicit correction is not enough and the user must actively recalibrate

What it is

Implicit correction is good at sliding a small translation. Once drift is large enough to have changed the nonlinear part—edge extrapolation has exploded, the homography no longer holds, glasses have slipped into a new mapping geometry—stacking more translation will fix the center and break the corners, or follow a bad label and collapse the mapping. Then the task has to stop and the user must actively recalibrate: look at known points again and refit. That is an upgrade of the correction method, not the error-budget management criterion itself; the budget answers “should we act,” this answers “a translation patch is no longer enough, replace the whole map.”

Why it happens

Online correction usually updates only low-order terms, because higher-order terms are unstable on sparse, noisy click labels. Once true error contains position-dependent bend (center still fine, worse toward the corners), low-order terms are not enough. A large headset slip, a different pair of glasses, walking from indoor light into sun, is a new set of extrinsics; the old higher-order coefficients are all void. Further implicit updates then crawl on the wrong manifold.

Active calibration costs an interruption and a bout of looking at points. It buys back a full-screen mapping including higher-order terms. The upgrade policy should be explainable: the system should be able to say “this bias is no longer something a nudge can fix,” not suddenly pop a nine-point chart.

Studying it

Make drift in two grades: pure translation versus position-dependent bend (for example extra offset only at the four corners). Contrast implicit translation only with forcing a full calibration once bend is detected. Outcomes: full-screen validation error, center-versus-corner difference, complaints about interruption. Signals of bend: click residuals that disagree in direction across regions; implicit updates that improve the center and worsen the corners. A lab that only places targets in the center will forever conclude implicit is enough, because the bend lives at the edge.

Where it stops holding

Some sessions are translation-only to the end, implicit can carry them, and forcing a full calibration is nagging. Conversely, a workout, glasses taken on and off, a child grabbing the headset, produce bend at any time and the implicit window is short. People who cannot finish a look-at-points task have no upgrade path; they need head direction or larger targets, not a calibration they cannot complete. Clinical records that require reportable accuracy may forbid riding implicit for the whole session from the start.

Applying it

• Watch whether residuals still look like a translation: when directions start fighting across regions, or implicit steps hit the cap and still do not fix, escalate to a full calibration.
• On escalate, say why in one sentence, and allow “enlarge targets and continue” as a temporary escape, not only the look-at-points path.
• Verify by adding bend at the corners and confirming the system does not keep translating the center into a faker fix, but enters active calibration once bend is detected.

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 · C8.10.3 Implicit recalibration can correct drift from interaction with known targets without interrupting the task
• Adjacent: C8.08 Eye-tracking calibration · C8.06 Foveated rendering
• Search: explicit recalibration · nonlinear gaze error · calibration session