Speaker Ted Adelson
Host Bill Freeman
Date recorded 14 July 2010
Objects are "things" but they are made of "stuff." In both human and machine vision, there has been much research on object recognition, but rather little on material perception. My lab is exploring material perception at various levels. At the low-level, we have identified certain image statistics, such as subband skewness, that are correlated with surface properties such as gloss. We suggest that there are mechanisms in early vision that are sensitive to such statistics, which could be easily computed with neural mechanisms. For our high-level studies, we are evaluating material recognition (categorization): the ability to look at something and decide whether it is, say, leather or plastic or cloth. We have assembled some image databases for use in exploring material recognition. In humans, material recognition can occur at high speed, even when the images are quite diverse. We have used machine vision and machine learning techniques in an attempt to achieve automated categorization of images in our database. While our system outperforms prior systems, it still falls far short of human performance.
©2010 Microsoft Corporation. All rights reserved.