 | ||
Illumination Statistics and Surface Reflectance Recognition | ||
Ron Dror | ||
MIT | ||
How can one distinguish between metal, plastic, and paper from a photograph? If one knew the amount of light incident on the surface from all directions, one could invert the computer graphics rendering process to determine reflectance properties such as shininess and gloss. If one does not know the illumination, however, the problem is underconstrained, because different combinations of illumination and reflectance could produce the same image. For example, a chrome sphere reflects the world around it, so if the illumination were just right, it could appear to be a ping-pong ball. Yet, in the real world, humans effortlessly recognize surfaces of different reflectance. We have found that the spatial structure of real-world illumination possesses a great deal of statistical regularity, akin to that described in the literature on natural image statistics. We argue that this regularity facilitates recognition of surface reflectance properties by both humans and machines. We have conducted psychophysical experiments suggesting that the human visual system makes use of real-world priors on illumination in recognizing surface reflectance properties. We have also designed a computer vision system to classify reflectance from a single monochrome image under unknown illumination. The system succeeds because of the predictable relationships between surface reflectance and certain statistics of the observed image. | ||