| Computer vision is hard because it attempts to understand a complicated scene with limited observation of one or many images. Many physical properties of a scene such as illumination, depth, surface orientation and reflectance are entangled and encoded into one single image. Measurement or separation of those physical properties can improve a very broad range of computer vision and graphics applications such as: more robust matching from illumination-invariant, reflectance-only images, more accurate shape from shading estimation from a reflectance-free image and image relighting by editing only the illumination component.;This dissertation proposes three techniques to measure and synthesize the illumination from one or many commonly available images of a scene. First, I present a technique to quickly gather images lit from different lighting positions, and synthesize new images by removing ambient light and only keeping directional illumination. Unlike previous expensive light stages, the technique remains low cost, yet can provide high quality basis images with very little ambient light for image relighting applications. Instead of using a controllable light stage, the second technique targets a single image from a class of commonly available outdoor image, and detects boundaries of abrupt illumination changes by carefully studying the physical properties of shadow boundaries to use as features in machine learning algorithms. This method showed substantial improvements when compared with previous shadow-detection methods on benchmark data sets. To conclude, I solve a more general problem: separating reflectance and illumination images from an single color image. I model the separation as a constrained optimization problem with a novel gradient-collinearity prior, and solve it with Gauss-Seidel method. The simple optimization scheme yields favorable results when compared with previous Retinex or machine learning algorithms. |
