Image statistics and the perception of surface reflectance
Author(s)
Sharan, Lavanya
DownloadFull printable version (33.85Mb)
Other Contributors
Massachusetts Institute of Technology. Dept. of Electrical Engineering and Computer Science.
Advisor
Edward H. Adelson.
Terms of use
Metadata
Show full item recordAbstract
Humans are surprisingly good at judging the reflectance of complex surfaces even when the surfaces are viewed in isolation, contrary to the Gelb effect. We argue that textural cues are important for this task. Traditional machine vision systems, on the other hand, are incapable of recognizing reflectance properties. Estimating the reflectance of a complex surface under unknown illumination from a single image is a hard problem. Recent work in reflectance recognition has shown that certain statistics measured o an image of a surface are diagnostic of reflectance. We consider opaque surfaces with medium scale structure and spatially homogeneous reflectance properties. For such surfaces, we find that statistics of intensity histograms and histograms of filtered outputs are indicative of the diffuse surface reflectance. We compare the performance of a learning algorithm that employs these image statistics to human performance in two psychophysical experiments. In the first experiment, observers classify images of complex surfaces according to the perceived reflectance. We find that the learning algorithm rivals human performance at the classification task. In the second experiment, we manipulate the statistics of images and ask observers to provide reflectance ratings. In this case, the learning algorithm performs similarly to human observers. These findings lead us to conclude that the image statistics capture perceptually relevant information.
Description
Thesis (S.M.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2005. MIT Institute Archives copy: p. 223 (last page) bound in reverse order. Includes bibliographical references (p. 217-223).
Date issued
2005Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer SciencePublisher
Massachusetts Institute of Technology
Keywords
Electrical Engineering and Computer Science.