Illumination Planning for Photometric Measurements

Abstract

Illumination is one of the basic factors that influence the quality of all intensity images in computer vision. Surprisingly, very few researchers have studied how illumination planning can make a difference in the quality of a computer vision method. Almost all intensity based computer vision techniques, including edge based methods, segmentation methods, and photometric methods, are affected by illumination. Since photometric methods directly model intensity, photometric measurements may be the intensity based computer vision technique which is most sensitive to illumination. Photometric techniques use physically based reflectance models to model image brightness base don lighting geometry, imaging geometry, local shape and reflectance. f we can control imaging geometry and illumination geometry, we can use a reflectance model in conjunction with measured image brightness to determine local shape and reflectance parameters. In photometric stereo, which is one class of photometric techniques, multiple images of an object are taken from the same viewpoint, using different light source directions, in order to determine local shape. The position of the light sources affects what parts of the object are illuminated and the accuracy with which you can recover the object's local shape. This thesis explores how illumination affects the reliability of the photometric measurement of surface orientation. We discuss the determination of surface orientation in the presence of intensity noise for the following: lambertian surfaces, rough diffuse surfaces, specular spike surfaces, and specular lobe surfaces. In addition to discussing how illumination affects the reliability of the photometric measurement of surface orientation, we also discuss how illumination affects the reliability of reflectance parameter measurements for specular lobe surfaces, and we discuss how to illuminate hybrid (specular lobe + lambertian) surfaces. There are two basic types of errors in photometric measurements: random errors (noise) and fixed errors. Random errors are due to the variance of the camera and digitizer. These are the errors that we try to predict with our planner. Fixed errors include: errors in light source direction, errors in light source radiance, and errors in the photometric function. Fixed errors can be accounted for by a careful calibration procedure.