Accuracy is the mean offset between estimated and true gaze

What it is

Spatial accuracy in eye tracking is the mean offset of estimated gaze from true gaze position, usually in degrees of visual angle. The person is truly looking at a point on the display; how far the system’s reports sit from that point on average is accuracy (larger offset, worse accuracy). It does not describe how samples scatter inside one fixation—that is precision—and it does not describe how calibration points were laid out.

Why it happens

True gaze direction is the visual axis; the instrument sees pupil and glint near the optical axis. Calibration aligns the two in the geometry of calibration time; leftover systematic difference after that—kappa not fully absorbed, unmodeled lenses, extrapolation off the validation points—shows up as an almost constant offset. Accuracy reports the size of that offset. Industry phrasing “accuracy of 0.5°” means about half a degree of mean offset, smaller better, which runs against the everyday “high accuracy”; read units and definitions in the paper.

Offset can have a direction: always right, always up. Mean offset is the magnitude of the vector average, or the mean of per-point offsets. A single scalar drops direction, and direction is exactly the cue for “the whole mapping translated” versus “only the edges are bad.”

Studying it

After calibration, have people fixate a set of held-out validation points; for each point take the mean estimate over a stable fixation, subtract the true coordinate, and average those offsets across points. Holmqvist’s methodology treats this as the minimum standard for reporting accuracy. Independent variables include validation location (center versus corners), head pose, glasses. Do not report accuracy on the calibration points themselves: that is fit residual and is optimistic. True points must be small, and the instruction must be “look at the center of the point,” or the truth itself is fuzzy and accuracy is inflated.

Where it stops holding

Without an external truth (free viewing of a film), accuracy cannot be measured directly; precision or task hits are proxies, and a proxy is not accuracy. With strabismus or binocular disagreement, “true gaze” depends on which eye; monocular accuracies must not be casually fused into one binocular number. A consumer headset measured only in the central 10° must not print 0.5° across the whole field. For low-vision users the “truth” may not be the geometric center of the point; their best optical image may sit slightly off.

Applying it

• When reporting accuracy inward or outward, state the validation set, head pose, glasses, and that the unit is visual angle.
• Accept on held-out points; if the center is good and the corners are bad, report by region, not one mean.
• Verify with the procedure above on the target device, and sit the mean offset next to smallest-target width rather than next to the brochure number.

Related

• Same group: C8.09.2 Precision is the spread of repeated measures; the two can vary independently · C8.09.3 High precision with low accuracy looks stable but systematically offset · C8.09.4 The lower bound on target size is set by accuracy, not precision
• Adjacent: C8.08 Eye-tracking calibration · C8.10 Gaze drift and recalibration
• Search: spatial accuracy · gaze offset · validation points