Three-dimensional object interpretation of monocular gray-scale images.

摘要

This thesis presents an implemented system for three dimensional object interpretation of monocular, gray-scale images. We successfully interpret scenes of composite objects in spite of self occlusion, surface markings, noise, and specularities. Test images include manufactured objects, with no prior restrictions on object position and orientation, or lighting.;Overall, the primary contribution of this work is the design and implementation of an end-to-end model-based interpretation system. Other contributions are the application of dynamically instantiated Bayesian networks to 3-D interpretation, a new geometric modelling system for vision, and Classics, a highly typed constraint system. The key to our success has been in combining new component technology into a single, coordinated system.;We interpret scenes using generic models from the bottom up, grouping features at one level into more sophisticated interpretations at the higher level. We use robust, existing theories to recover scene (3-D) information from the image (2-D), and obtain only a small set of reasonable 3-D hypotheses. This avoids the combinatorics of comparing object view models with tuples of image features. We automatically derive Bayesian recognition networks from generic object and feature models at all levels. The strongest evidence leads to the instantiation of a few model hypotheses, which are then used to predict image features. Weaker evidence found by prediction gives support or denial to the hypotheses. This incorporates a maximum of evidence with a minimum of misleading conclusions.;Bayesian networks of increasing complexity are instantiated in real time to calculate probabilistic support for hypotheses. Probabilistic reasoning focuses attention on most likely interpretations first.;We develop a generic object modeler based on a highly typed constraint system, also of our own creation. Models are defined using geometric, algebraic and other constraints with which automated reasoning is possible.;The current implementation uses image intensity edge information as input. The method is generalizable to using shading, stereo, color, range and other data.