Alvey MMI-007 Vehicle Exemplar: The Knowledge Based Approach

Since the seminal work of Roberts (1965), sceneunderstanding research has continued to be dominated by two main themes: the measurement of local attributes of the image to identify features which characterise objects in the scene, and the use of prior expectations, often explicitly expressed as models, to guide the interrogation of the image. The two themes correspond to the distinction between data-driven methods, and goal-directed methods encountered in many studies in AI. Data-driven methods may be used to express lowlevel knowledge about image properties, but the goals of vision involve knowledge of objects and their relationships in the scene which demands high-level representations with little direct correspondence in the image data. The central task confronting practical vision systems is to resolve the distinction, and interpret the image data in terms of object abstractions.