Abstract

The problem of assigning labels from a fixed set to each member of a set of sites appears at all levels of computer vision. Recently, an optimization algorithm known as Highest Confidence First (HCF) [Chou, 1988] has been applied to labeling tasks in low-level vision. Examples of such tasks include edge detection, in which each inter-pixel site must be labeled as either edge or non-edge, and the integration of intensity and sparse depth data for the labeling of depth discontinuities and the generation of dense depth estimates. In these tasks, it often outperforms conventional optimization techniques such as simulated annealing[Geman and Geman, 1984], Monte Carlo sampling[Marroquin, 1985], and Iterative Conditional Modes (ICM) estimation[Besag, 1986].

© 1989 Optical Society of America