Fixation count prediction for textural scenes

Abstract

The human eye collects visual information by means of saccades and fixations. Recent work shows that fixation locations are not arbitrary. On the contrary, they tend to cluster on the salient regions of the scene. Automatic estimation of the number of fixations on an image has uses in many applications and contexts including computer vision (e.g., robot vision, compression, salience estimation) and human-computer interaction (e.g interface usability assessment). In this study, we present an algorithm for estimating the number of fixations on parts of an image based on local descriptors using supervised regression models on the DOVES eye movements dataset. Our results suggest that in the absence of contextual information, local descriptors can be used to generate a reasonably accurate fixation intensity map of an image.