Bischof, W. F. and Caelli, T. (1997). SURE: Scene understanding by rule evaluation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19,1284-1288.
We consider how machine learning can be used to help solve the problem of identifying objects or structures composed of parts in complex scenes. We first discuss an automatic conditional rule generation technique (CRG) that is designed to describe structures via part attributes and their relations. It does so by automatically "pre-compiling" part-indexed decision trees where the branches define the types of pattern structures necessary to identify and to generalize from the different training examples. We then show how the resultant rules can be used for region labeling and examine grouping and constraint propagation techniques that are required for the identification of objects in complex scenes.
Back to publications.