Drift is the gradual offset of gaze estimates from true position after calibration

What it is

Drift is the process, after calibration is already done, in which the gaze estimate gradually leaves true position. It is error that crawls, not a bad fit at calibration time (that was inaccurate from the start), and not a step from one push of the glasses (that is broken conditions). Drift makes the same coefficients point somewhere else ten minutes later; users report “it was fine a moment ago, now it is always a bit off.”

Why it happens

The mapping freezes optical geometry at calibration into coefficients. Afterward slow variables keep moving: the pupil scales with load and display brightness, the tear film dries, the head sinks in the chair, a headset strap eases a millimeter. Each term is small; together they show up as accuracy worsening over minutes. Because it is slow, precision inside a single fixation can stay good—the cursor is steady, just steadily in the wrong place.

Drift is not a random walk; it often has a dominant direction tied to gravity, strap elasticity, and the display’s brightness curve. It therefore looks like high-precision/low-accuracy unfolded in time. Without a time axis, reporting only accuracy at the end of calibration writes drift as “this instrument is inaccurate.”

Studying it

Measure a validation set immediately after calibration, then remeasure the same points every few minutes without recalibrating, and plot mean offset against time. Log lighting and task load; they drive the pupil arm. Contrast a chin rest with free sitting to separate slow head sink from purely ocular optics. A ten-minute laboratory curve does not extrapolate to a two-hour film; remeasure across a session of the target length. Reports of drift must state starting accuracy, or the climb is invisible.

Where it stops holding

Some devices barely climb in the minutes after calibration; repeating “it will drift” copies another instrument’s disease. Conversely, a remote tracker plus a loose sit can blow the budget inside one short task. In children and people who fidget, the head’s “slow” change is fast, and the process no longer matches “gradually.” Fully calibration-free models that adapt every frame have no “after calibration” origin; their drift has to be rewritten as adapter lag, not this definition.

Applying it

• Treat accuracy at the end of calibration and accuracy mid-session as two numbers; do not accept on boot only.
• For sessions longer than the target duration, reserve a mid-session remeasure rather than assuming zero drift.
• Verify by remeasuring the same validation points in the target posture and duration, plotting offset against time, and seeing when it crosses the error budget.

Related

• Same group: 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 · C8.10.4 Beyond a point, implicit correction is not enough and the user must actively recalibrate
• Adjacent: C8.08 Eye-tracking calibration · C8.09 Precision versus accuracy
• Search: gaze drift · post-calibration · temporal offset