Precision is the spread of repeated measures; the two can vary independently

What it is

Spatial precision in eye tracking is the scatter among repeated samples under the same fixation, often RMS or standard deviation, also in visual angle. It asks whether the needle jitters, not whether the needle points at the truth. Accuracy can be poor (the whole cloud sits 1° right of truth) while precision is good (the cloud scatters only 0.1°); the reverse is possible too. They vary independently; neither substitutes for the other.

Why it happens

Scatter comes from sensor noise, jitter in pupil-edge extraction, microsaccades, and physiological tremor. The first two are the instrument; the last two are the eye. Higher sample rate, better edge fits, and modest smoothing can suppress instrument noise; microsaccades cannot be suppressed, and suppressing them too hard filters real saccades too. When precision is poor, even if the mean happens to sit on the target, single-frame coordinates hop in and out, and dwell clocks and hit tests chatter.

Independence is the error split: the systematic part goes to accuracy, the random part to precision. Calibration mostly moves the systematic part and helps RMS little; a more stable IR lamp or a cleaner pupil ellipse mostly improves precision while mean offset may not move at all. Collapsing both numbers into one “precision of 0.5°” is a common industry mix-up; when reading, check whether they measured offset or scatter.

Studying it

Have someone hold a point (or use an artificial eye to remove microsaccades), take a stretch of samples labeled as a fixation, and report RMS or inter-sample SD. Holmqvist recommends reporting accuracy and precision together and stating the filter. Contrasting a real eye with an artificial eye separates physiological jitter from instrument noise. Independent variables: lighting, lenses, smoothing window. Smoothing makes the precision number prettier and delays real motion; it must be reported with latency. Do not compute RMS on saccade stretches; that is motion, not precision.

Where it stops holding

Precision on an artificial eye is an instrument ceiling; live people always scatter more, and a product cannot use the artificial-eye number as user precision. Uncompensated head motion writes head jitter into “gaze precision.” Epilepsy, nystagmus, and severe fatigue collapse precision; no calibration will pin the landing. High precision does not license small buttons: the mean can still sit outside the button.

Applying it

• Record and display accuracy (validation-point offset) and precision (fixation-segment RMS) separately; do not mint one ambiguous “precision.”
• Smooth only to suppress instrument noise, and publish window length and the effect on saccade detection.
• Verify under target lighting and target glasses, measuring RMS on a real eye and (if available) an artificial eye, confirming that dwell and hit logic do not chatter in and out at that RMS.

Related

• Same group: C8.09.1 Accuracy is the mean offset between estimated and true gaze · 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: C8.01 Fixation and saccade · C8.08 Eye-tracking calibration
• Search: spatial precision · RMS · gaze jitter