|
|
|
|
Evans & Sutherland Distinguished Lecture Series Image Statistics and Surface Perception Edward Adelson Massachusetts Institute of Technology Host: Bill Thompson Abstract When we look at the world, we see objects made of materials. In human and machine vision, many researchers study object perception, and only a few study materials. On the other hand, people in computer graphics are obsessed with materials, because they know how important they are to human viewers. One reason that material perception is under-studied is that it is just plain hard. Understanding the statistics of natural scenes turns out to be crucial. Some basic concepts can be traced back to the work of Stockham and colleagues, who sought to separate illumination from reflectance by characterizing their image statistics in the frequency domain. Edwin Land developed a gradient based approach in his Retinex models. We are reconsidering this problem using techniques from machine learning and multiscale image decomposition (e.g., wavelets). We are also probing the human visual system with illusions to uncover some of the principles it uses in evaluating reflectance. Another important area is the perception of gloss. Here we find that scene statistics, in the form of the illumination distributions surrounding a surface, are absolutely critical, and are implicitly known by human observers. For some simple surface geometries, we have devised a machine vision system that can classify surfaces of varying lightness and gloss when they are viewed in natural illumination. With unnatural illumination, both the machine and humans get it wrong.
|
School of Computing • 50 S. Central Campus Dr. Rm. 3190 • Salt Lake City, UT 84112
801-581-8224 • Send comments to webmaster@cs.utah.edu
Disclaimer