Caelli, T. and Bischof, W. F. (1995). Finding complex targets in complex scenes using machine learning: A viable surveillance paradigm? Proceedings of the workshop on Artificial Intelligence in Defence, Canberra, Nov 14, pp. 53-71.

In this paper, we discuss automatic rule generation techniques for learning relational properties of 2-D visual patterns and 3-D objects from training samples where the observed feature values are continuous. In particular, we explore a new conditional rule generation method that defines patterns (or objects) in terms of ordered lists of bounds on unary (single part) and binary (part relation) features. The technique, termed Conditional Rule Generation (CRG), was specifically developed to integrate the relational structures of graph representations of patterns and the generalization characteristics of Evidenced-based Systems (EBS). CRG takes into account the label-compatibilities that should occur between unary and binary rules in their very generation, a condition that is, generally, not guaranteed in well-known Rule Generation and Machine Learning techniques as they have been applied to problems in Computer Vision. We show how this technique applies to the recognition of complex targets and of objects in scenes, and we show the extent to which the learned rules can identify patterns and objects that have undergone non-rigid distortions.

Back to publications.