Sub-centimeter 3D gaze vector accuracy on real-world tasks: an investigation of eye and motion capture calibration routines

Abstract

Measuring where people look in real-world tasks has never been easier but analyzing the resulting data remains laborious. One solution integrates head-mounted eye tracking with motion capture but no best practice exists regarding what calibration data to collect. Here, we compared four ~1 min calibration routines used to train linear regression gaze vector models and examined how the coordinate system, eye data used and location of fixation changed gaze vector accuracy on three trial types: calibration, validation (static fixation to task relevant locations), and task (naturally occurring fixations during object interaction). Impressively, predicted gaze vectors show ~1 cm of error when looking straight ahead toward objects during natural arms-length interaction. This result was achieved predicting fixations in a Spherical coordinate frame, from the best monocular data, and, surprisingly, depends little on the calibration routine.