Computing intrinsic images

摘要: Low-level modern computer vision is not domain dependent, but concentrates on problems that correspond to identifiable modules in the human visual system. Several theories have been proposed literature for computation of shape from shading, texture, retinal motion spatiotemporal derivatives image intensity function, and like. The with existing approach are basically following: (1) The employed assumptions very strong (they present a large subset real images), so most algorithms fail when applied images. (2) Usually constraints geometry physics problem enough guarantee uniqueness computed parameters. In this case, additional about world used, order restrict space all solutions unique value. (3) Even if no at used physical parameters, then cases resulting robust, sense there slight error input (i.e. small amount noise image), results catastrophic output (computed parameters). It turns out several available cues combined, above-mentioned disappear; compute uniquely robustly intrinsic parameters (shape, depth, motion, etc.). In thesis machine explored its basics. A low level mathematical theory presented robust computational aspect envisages cooperative highly parallel implementation, bringing information five different sources (shading, contour stereo), resolve ambiguities ensure stability shading analysis analyzed detail.