Learning continuous models for estimating intrinsic component images

Abstract

The goal of computer vision is to use an image to recover the characteristics of a scene, such as its shape or illumination. This is difficult because an image is the mixture of multiple characteristics. For example, an edge in an image could be caused by either an edge on a surface or a change in the surface's color. Distinguishing the effects of different scene characteristics is an important step towards high-level analysis of an image. This thesis describes how to use machine learning to build a system that recovers different characteristics of the scene from a single, gray-scale image of the scene. The goal of the system is to use the observed image to recover images, referred to as Intrinsic Component Images, that represent the scene's characteristics. The development of the system is focused on estimating two important characteristics of a scene, its shading and reflectance, from a single image. From the observed image, the system estimates a shading image, which captures the interaction of the illumination and shape of the scene pictured, and an albedo image, which represents how the surfaces in the image reflect light. Measured both qualitatively and quantitatively, this system produces state-of-the-art estimates of shading and albedo images.

(cont.) This system is also flexible enough to be used for the separate problem of removing noise from an image. Building this system requires algorithms for continuous regression and learning the parameters of a Conditionally Gaussian Markov Random Field. Unlike previous work, this system is trained using real-world surfaces with ground-truth shading and albedo images. The learning algorithms are designed to accommodate the large amount of data in this training set.