Representing and Using Knowledge of the Visual World
Methodology for the representation of knowledge is a fundamental aspect of research in Computational Vision. The properties of objects and the relationships among objects must be represented for a given task domain and the representation must also support efficient processes of recognition and search. These twin criteria for evaluating knowledge representations are called descriptive adequacy and procedural adequacy respectively and are applicable to both early visual processing and high-level visual recognition. All vision requires knowledge representations which exhibit both descriptive and procedural adequacy. Here, we examine the knowledge representations which have been used in high level vision. In particular, well-understood network consistency representations are shown to have a number of inherent limitations. We propose the use of schemata as a unifying representational formalism. Mapsee2, a recent experimental system using this representation, is used to illustrate the advantages of schema-based techniques.
KeywordsProcedural Knowledge Computational Vision Visual World Procedural Adequacy Schema Instance
Unable to display preview. Download preview PDF.