RealEye Webcam v3 Accuracy Across Skin Tones

Published: 3 min read

RealEye Webcam v3 delivers usable eye-tracking accuracy for every skin-tone group in our validation data. Accuracy is not identical across groups — and we report the full, honest spread below, including the group with the largest error.

This article covers RealEye Webcam v3, the default webcam eye-tracker for new RealEye studies. For the overall accuracy page (computers, all trackers), see RealEye Accuracy on Computers.

#The headline numbers

The figures below come from ETv3 validation data collected from 5,375 real webcam eye-tracking sessions: 913,699 webcam images and 708,272 valid samples after outlier removal.

  • Overall ETv3 accuracy: median error 6.7% of the screen diagonal, mean error 8.0%

Accuracy here means how far the predicted gaze point is from the actual point the participant was looking at, measured as a percentage of the screen diagonal. Lower is more accurate. The median is the typical error (half of all samples were better), while the mean is the average.

#Accuracy by skin tone

The chart and table below show median and mean ETv3 error after outliers are removed, for six apparent skin-tone groups: White, East Asian, South Asian, Latino, Other, and Black. Every group stays below 10% median error, and the spread across groups is modest — from 5.3% (White) to 9.1% (Black) median error.

Chart: ETv3 median gaze error by apparent skin tone

Table: ETv3 accuracy by apparent skin tone

#What this means in practice

  • ETv3 works across diverse faces. More than half of the sessions in this dataset — 55% — come from non-White groups, so these results are not based on a narrow sample.
  • Accuracy is good for every group. In all six groups, the median gaze error is under 10% of the screen diagonal, which is well suited to heatmaps, Areas of Interest (AOIs), and attention analysis.
  • The differences between groups are real, but modest. The gap between the lowest and highest group medians is 3.8 percentage points of the screen diagonal. Both ends of that range produce reliable eye-tracking results for typical studies.
  • For studies that are especially sensitive to accuracy, follow the same best practices as always: good lighting, a quality webcam, participants seated at a comfortable distance from the screen, and appropriately sized AOIs. See RealEye Accuracy on Computers and How to get accurate results.

#A note on how these groups were defined

  • The skin-tone groups are machine-classified apparent skin tone, detected automatically from each participant's first webcam image. They are not self-reported ethnicity — they describe how faces appear to the computer vision model, not how participants identify themselves.
  • These figures apply to RealEye Webcam v3 only. They do not describe RealEye Webcam v1, smartphones, or the overall accuracy figures on other pages.
  • Outliers were removed before calculating these figures; the numbers above reflect the remaining 708,272 valid samples.
  • 302 sessions (5.6%) had an unknown skin tone. They are included only in the overall figures, not in any group.

#Related articles

💌 If you have any further questions, feel free to reach out to our team at support@realeye.io