Another failing of traditional vision systems is the lack of attention given to purposeful, selective control of the processing. This is again a key issue when real-time, dynamic applications are being developed. Ballard's influential paper [2] signalled the start of a new consensus in the computer vision community that vision is active, highly selective and purposeful. Rao and Ballard [49] have also proposed an active vision architecture, inspired by the organisation of human visual processing, that uses simply acquired and indexable iconic representations. Interdisciplinary research also shows that the task and the nature of the scene determine visual attention and can allow selective tuning of visual processing [68]. This requirement for highly selective visual processing was the main theme of VAP [19,20], which was a major European project to develop active visual processing. Much of this active vision research has concentrated on camera control, navigation and lower-level visual tasks which do not involve visual understanding using stored knowledge. However, ideas from active vision have been extended and applied to high-level reasoning [13,33]. This control of visual processing is deeply task-dependent and usually requires indexable knowledge structures for real-time systems.