We are exploring both low-level and high-level aspects of material perception. At the low-level, we ask: are there simple image statistics that have an impact on material appearance? We find that modifying subband statistics -- especially subband skewness -- can change apparent gloss and other surface qualities. We suggest that there are mechanisms in early vision that are sensitive to such skewness, 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 the basic properties of material recognition. We feel it is a distinct skill, which is different from object recognition and different from texture or color perception. Material recognition can occur at high speed. We find that humans can categorize a material almost as quickly as they can make a simple color judgment. We are looking for image statistics that will allow a machine vision system to do material categorization, with limited success so far.
展开▼