Learning Object Models From Visual Observation and Background Knowledge.

Abstract : This research project aims to use machine learning techniques to improve the performance of three dimensional vision systems. Building on our earlier work, our approach represents and organizes models of object classes in a hierarchy of probabilistic concepts, and it uses Bayesian inference methods to focus attention, recognize objects in images, and make predictions about occluded parts. The learning process involves not only updating of the probabilistic descriptions in the concept hierarchy but also involves changes in the structure of memory, including the creation of novel categories, the merging of similar classes, and the elimination of unnecessary ones. An evaluation metric based on probability theory guides decisions about such structural changes, and background knowledge about function and generic object classes further constrains the learning process. We plan to carry out systematic experiments to determine the ability of this approach to improve both classification accuracy and predictive ability on novel images.