A multi-level geometric reasoning system for vision

摘要

Geometry is known to play an important role in image understanding and machine vision. Geometric reasoning is employed in different forms in model formation and model matching problems in model-based vision research. Geometric representations and reasoning methods are often used implicitly in the design of vision systems and algorithms used there. An approach towards model-based vision in which geometric and algebraic reasoning is explicit is discussed. A multi-level reasoning system for machine vision based on this approach is being designed and developed at General Electric Corporate Research and Development. Three key components of this system—a hierarchical organization of geometric knowledge for its systematic and efficient use as well as to control the search space, labeling algorithms and algebraic reasoning algorithms—are presented. The use of such a system for deriving a three-dimensional model from two-dimensional images as well as for using a two-dimensional image for matching against another two-dimensional image (called view consistency problems) is discussed.