Probabilistic Learning of Three-Dimensional Object Models
Gregory M. Provan, Thomas O. Binford · 2011
In this paper we report on an approach to learning object models for use in recognition and reconstruction. Our framework represents objects in an image using generalized cylinders and organizes knowledge about classes of objects in a Bayesian network. The recognition process involves propagating evidence through this inference network, whereas learning relies on updating of the network 's conditional probabilities based on training cases. We report preliminary experimental results with synthetic data that suggest our method improves its recognition accuracy with experience. We also consider our framework's relation to other research on learning object knowledge for image understanding. 1. Introduction The image-understanding process relies on accurate knowledge. This statement holds for all levels of visual processing, but seems especially true at the later stages, where object recognition and reconstruction require models of objects or object classes that occur in the domain. And th...