Learning Image Annotation: The CITE System

Craig Dillon
Department of Computer Science
Curtin University of Technology

Terry Caelli
Center for Mapping
The Ohio State University



Abstract

In this paper we consider how more-recent machine learning techniques can be used in building adaptive and trainable image annotation systems that can mimic human performance. In this case the annotation is defined with respect to domain-specific labels for image regions, hierarchies, and, in general, cliques that are of specific importance to the application domain. To illustrate these issues, we consider one of our more recent annotation systems, the CITE system, where the domain knowledge consists of hierarchies of labeled and attributed graphs, and how it has been used to annotate outdoor, airport, and office scenes consisting of different types of scenarios and information.

Keywords: image labeling, image animation, learning, knowledge-directed computer vision, image segmentation