Abstract

L. Itti, P. F. Baldi, The salience map: a local surprise detector?, Cognitive Neuroscience Annual Meeting, Workshop on the Saliency Map, New York, N.Y., Apr 2005.

Abstract: To investigate the extent to which salience maps may highlight regions in the visual field with high information content, we propose a subjective definition of information we call surprise, to quantify how data affects a natural or artificial observer, by measuring the difference between prior and posterior beliefs of that observer. We argue that surprise is better suited to studying subjective aspects of brain function and behavior than Shannon information, particularly sensory processing and novelty detection. Thus, we build a computational model of early vision and attention, which topographically computes visual surprise at every location in a salience map. It outperforms Shannon information and other models in predicting gaze of four humans watching 50 complex videoclips. This suggests that visual locations which appear as salient to an observer may do so because they are surprising more than because they are informative.

Themes: Computational Modeling, Model of Bottom-Up Saliency-Based Visual Attention, Computer Vision, Bayesian Theory of Surprise, Human Eye-Tracking Research