UBC Theses and Dissertations
Learning to recognize objects in images : acquiring and using probabilistic models of appearance Pope, Arthur R.
We present a method of recognizing three-dimensional objects in intensity images of cluttered scenes, using models learned from training images. By modeling each object with a series of views, representing each view with a large variety of features, and describing each feature probabilistically, our method can learn to recognize objects of unusual complexity. An object is modeled by a probability distribution describing the range of possible variation in the object's appearance. This distribution is organized on two levels. Large variations are handled by partitioning training images into clusters corresponding to distinctly different views of the object. Within each cluster, smaller variations are represented by distributions characterizing uncertainty in the presence, position, and measurements of various discrete features of appearance. Many types of features may be used, ranging in abstraction from segments of intensity edges to perceptual groupings and regions. Our learning method combines two activities: (a) an incremental conceptual clustering algorithm identifies groups of training images corresponding to distinct views of the object; (b) a generalization algorithm consolidates each cluster to produce a summary description characterizing its most prominent features. The method produces models that are both economical and representative, balancing these competing goals through an application of the minimum description length principle. Recognition requires matching features of a model with those of an image. Our matching method is a form of alignment: from hypothesized feature pairings it estimates a viewpoint transformation; from that it finds additional pairings, and so on. The method uses constraints based on features' positions, numeric measurements, and relations, assigning to each an importance commensurate with its uncertainty as recorded by the model. Decisions are ordered so that more certain features are paired sooner, while ambiguous choices are postponed for later resolution, when additional constraints may be known. We also describe a system implemented to test our recognition learning method. Experiments demonstrate the system's performance at tasks including learning to recognize complex objects in cluttered scenes, and learning to discriminate two quite similar objects.
Item Citations and Data