Abstract
When searching for a target among distractors humans often make saccadic eye movements every 200–300 ms. There have been many attempts to model these eye movements in a way that accurately reflects the observed human behavior (Najemnik & Geisler, 2005, Nature; Rao et al., 2002, Vision Research; Beutter et al., 2003; JOSA A). However, there are no standard methods to compare the eye movement fixations of human and models. Here, we propose a maximum likelihood metric that quantifies the probability of observing a human saccade or a sequence of saccades given that a model is driving the saccades. This probability is calculated by taking into account both the saccade noise inherent to the human brain as well as the error stemming from the particular eye tracking system used. We compare the likelihood metric to other common metrics including: a) the distance between the measured human and predicted model saccade endpoints (distance metric), b) a correlation metric between the human and model spatial cluster of fixations (cluster correlation metric), and c) the percentage of the trials in which the eye movements were directed towards the target for the models and human observers (percentage correct metric). We highlight theoretical situations when the distance, cluster-correlation, and percentage correct metrics fail to adequately distinguish between models. We finally apply the common metrics and the new likelihood metric to actual measured human saccades in a series of 400ms search tasks for a variety of targets with contrast noise (Gaussians: full width at half of maximum (FWHM) = 0.384 degrees, single frame (25ms) signal to noise ratio (SNR) = 2.58 or Gabors: FWHM of Gaussian envelope = 0.384 degrees, spatial frequency = 9.8 cycles/degree, single frame (25ms) SNR = 7.41).
National Science Foundation (BCS-0135118).