Abstract


= PDF Reprint,     = BibTeX entry,     = Online Abstract


Click to download BibTeX data Clik to view abstract L. Itti, P. F. Baldi, A surprise theory of attention, In: Proc. Vision Science Society Annual Meeting (VSS05), May 2005. (Cited by 12)

Abstract: [ORAL] Attention in biological and artificial systems rapidly selects important information from within massive sensory inputs, a process key to survival. When time lacks for detailed sensory analysis, finding important information must rely on heuristic or approximate computations. To quantitatively characterize these computations, we propose a Bayesian definition of important information we call surprise. Surprise quantifies how data affects a natural or artificial observer, by measuring the difference between prior and posterior beliefs of the observer. We argue that surprise subsumes and extends previous often ad-hoc notions of stimulus saliency and novelty, casting them into a single theoretical framework derived from first principles. To test this, we measure the extent to which ten image-based metrics that highlight different facets of important information may predict gaze recordings of four human observers watching 50 complex videoclip stimuli, including television broadcast and video games (about 30 minutes in total). At the target location of each of the 10,192 saccadic gaze shifts recorded, compared to at random locations, we evaluate intrinsic visual properties of the video clips using the ten computational metrics, including a surprise metric. Extending previous findings, but for dynamic scenes, we find that humans preferentially gaze towards locations where local entropy, contrast, information, color, intensity and orientation responses are higher than expected by chance (sign tests, p<1.0E-100 or better). Furthermore, metrics computing dynamic image features like flicker, motion, saliency and surprise correlate even better with human eye movements. Out of all metrics, surprise significantly stands out as best-scoring (t-tests, p<1.0E-100 or better). Our data shows that guiding attention towards intrinsically surprising stimuli is an efficient shortcut to important information.

Note: Oral presentation

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

 

Copyright © 2000-2007 by the University of Southern California, iLab and Prof. Laurent Itti.
This page generated by bibTOhtml on Wed Feb 15 12:13:56 PST 2017