J CMU Image Understanding Program Take 0 Kanade

Image Understanding tesearch at CMU addresses a bmad spectrum of issues,from applications of machine vision to the development of new sensors and basic vision science. The Rapid site modeling-the generation of a detailed three-dimensional model of a surveyed site-is a critical problem for both military and civilian applications, including geographic database generation for cartography, reconnaissance , damage assessment, combat simulation, aod autonomous air/gmund vehicle navigation. Central to the site modeling problem is the develop ment of efficient and reliable Image Understanding techniques to analyze and extract precise threedmensional shape information from multiple visual or range images taken from a moving platform, such as a scouting groundair vechicle, or a stereoscopic camera. We have recently developed three new techniques: 1) the factor-ization method for shape and motion recovery from an image sequence; 2) the multi-baseline stereo method for reliable and dense depth mapping; and 3) the landmark object modeling method from range image sequence. AU of these methods have been tested with images taken in a controlled laboratory environment (to provide ground-truth data for quantitative evaluation) and taken outdoors under real lighting and geometry conditions. Their performance has been demonstrated to exceed that of previous methods. The structure from motion problem-recovering scene geometry and camera motion from a sequence of images-has attracted much of the attention of the vision community over the last decade, and yet it is common knowledge that existing solutions work well for perfect images, but are very sensitive to noise "%ere are two fundamental reasons for this. First, when camera motion is small, effects of camera rotation and translation arc conjugate: for example , rotation about the z-axis and translation along the x-axis both generate a very similar change in an image. Any attempt to recover or differentiate between these two motions is naturally noise sensitive. Second, computation of shape as relative depth, for example, the height of a building as the difference of depths between the top and the bottom, is very sensitive to noise, since it is a small difference between large values. These difficulties are especially magnified when the objects ~IE distant from the camera relative to their sizes, which is usually the case for interesting applications such as site modeling. Recently we (Tomasi and b a d e) [11[21 observed that both difficulties disappear when the problem is reformu-lated in worldcentend coordinates unlike the conventional cameracentered formulation. This new formulation links objectcentered shape …