| In this thesis we explore the problem of bottom-up figure-ground discrimination. Given a single image, we are interested in partitioning the image into a small number of distinct objects using only generic image cues. We address this problem from the point of view of two broad fields of computational vision: image segmentation, and perceptual grouping. Within the field of image segmentation, we present a novel algorithm that uses spectral embedding and min-cut to generate high-quality segmentations of arbitrary images. We describe the mathematical foundation of spectral embedding, show a relationship between spectral embedding and anisotropic smoothing, and examine the generation of seed regions for min-cut. We then present a quantitative study of segmentation quality on the Berkeley Segmentation Database. We propose suitable measures of segmentation quality, and show comparative segmentation quality results for our algorithm and three competing segmentation methods. Within the field of perceptual grouping, we present a general search-based framework for contour extraction on cluttered images. The algorithm is based on locally normalized affinities for search control, and supports cue integration for increased robustness. We develop the algorithm in two parts, first, we demonstrate the robustness and efficiency achieved by our search framework in the context of convex group detection, and then describe suitable constraints that can be used to find non-convex contours. We discuss the problem of evaluating contour saliency, and demonstrate the use of colour information to increase robustness. Finally, we discuss directions for future research for both of the components of our work. |
