Calibration accuracy decays when head pose, glasses, or lighting change after calibration

What it is

Calibration fits geometry “as it was.” Afterward the user changes head pose, slides glasses up the nose, or walks to a window, and the pupil–reflection relation in the camera is no longer the one that was fit; mapping accuracy drops. This is broken calibration conditions, not the kind of drift in which coefficients crawl on their own over time—that is a continuous process. Decay here can be a step completed by one push of the glasses.

Why it happens

Head pose changes the eye’s distance and angle to the camera, so the image vector for the same gaze direction changes. Sliding glasses move both the lens optical axis and frame reflections; the corneal glint may jump to another hotspot on the lens. Lighting changes pupil diameter: the pupil as a near-planar disc, when it changes size, moves the estimated pupil center in 3D even if the corneal reflection stays put. All three can happen seconds after calibration; “having used it a while” is not required.

Remote desktop trackers are most pose-sensitive because the camera is fixed and the head moves in a large box. Wearables nail the cameras to the face, so pose hurts less, but glasses slipping on the headset and a pupil step from indoor to outdoor still hit the mapping directly. Makeup, tears, and fog belong to the same class: optical conditions changed, geometry at the calibration points is void.

Studying it

From a just-calibrated baseline, introduce three disturbances separately: a prescribed head yaw/pitch, moving the glasses several millimeters along the nose, toggling a window light or display brightness; then measure the angular-error increment on the same validation points. Report the step size; do not only report “after ten minutes”—ten minutes may have changed no condition. Stratify frames, contacts, and bare eye. Contacts slip less, but a drying lens wrinkle is another optical change. Constant laboratory lighting erases the illumination arm.

Where it stops holding

On a wearable, with no glasses and stable lighting, pose hurts the mapping less than on a remote tracker, and this decay is milder. Conversely, in a car, at a window desk, or with glasses taken on and off, calibration is valid mostly in the current pose. Some systems deliberately allow small head motion during calibration to absorb pose into a more robust model; decay slows, but fixation quality on the points is harder to guarantee. Pathological anisocoria or drug-induced miosis can break the lighting→pupil→mapping chain with no environmental change.

Applying it

• Tell people in the calibration prompt to hold the head pose and glasses position they will actually use, and not to calibrate in a dark room and then open a window.
• When a clear condition step is detected (head leaving the calibration box, a pupil-diameter jump, a changed lens hotspot), signal that the mapping may be invalid rather than keep pointing with the old coefficients.
• Verify on the target device by pushing glasses, turning the head, and changing display brightness, and measure whether the error step already exceeds half the smallest target.

Related

• Same group: C8.08.1 Calibration has the user look at known screen points to build an individual mapping · C8.08.3 Whether to recalibrate depends on whether drift exceeds the acceptable error · 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: calibration degradation · head pose · spectacle slippage