Item

Abstract

In the computer vision community, so-calledinterest operatorshave become verypopular in recent years. An interest operator is a function that, given an image, re-turns a set ofinterestingpoints that can then be used to e.g. categorize images or todetect objects within images. To our knowledge, all of the widely used interest opera-tors are based onheuristicmeasures of local interestingness (e.g. contrast, second orderstructure, entropy, etc.). While this works well in numerous image processing applica-tions, it seems natural to ask how (if at all) they relate to the visual system’s notion ofinterestingness. This has been explored by e.g. Privitera and Stark [1], who found sig-nificant correlations between fixation points and locations selected by existing heuristicinterest operators. Reinagel and Zador [2] took the opposite approach and reportedthat patches around fixation points tend to have higher contrast and that point-wisecorrelations decay with eccentricity.In this work, we want to find regularities in the image structure around fixationpoints. Thus, we follow the spirit of [2], but instead of verifyingheuristichypothesessuch as ”higher contrast” or ”correlations decay with eccentricity”, our goal is toinferhypothesesfrom the data, using methods from statistical learning. To this end, we traina non-linear classifier on fixated vs. randomly selected image patches, i.e. we learn aninterest operator from human eye movements.At present, we are exploring various ways to deal with two major problems inherentin our approach. First, when a fixation point is recorded a some particular position,the size of the attracting image region around that point is unknown. Second, fixationpositions are subject to ”inaccurate” gaze positioning and measurement noise, which issubstantial compared to pixel sizes. At the workshop, we will present first experimentalresults showing that statistical methods are capable of learning discriminative imagefeatures around fixation points.