Generative Object Definition and Semantic Recognition

Abstract

''What is the difference between a cup and a door?'' These kinds of questions have to be answered in the context of digital libraries. This semantic information, which describes an object on a high, abstract level, is needed in order to provide digital library services such as indexing, markup and retrieval. In this paper we present a new approach to encode and to extract such semantic information. We use generative modeling techniques to describe a class of objects: each class is represented by one algorithm; and each object is one set of high-level parameters, which reproduces the object if passed to the algorithm. Furthermore, the algorithm is annotated with semantic information, i.e. a human-readable description of the object class it represents. We use such an object description to recognize objects in real-world data e.g. laser scans. Using an algorithmic object description, we are able to identify 3D subparts, which can be described and generated by the algorithm. Furthermore, we can determine the needed input parameters. In this way, we can classify objects, recognize them semantically and we can determine their parameters (cup's height, radius, etc.).