To recognise a previously seen object, the visual system must overcome the variability in the object's appearance caused by factors such as illumination and pose. It is possible to counter the influence of these factors, by learning to interpolate between stored views of the target object, taken under representative combinations of viewing conditions. Routine visual tasks, however, typically require not so much recognition as categorisation, that is making sense of objects not seen before. Despite persistent practical difficulties, theorists in computer vision and visual perception traditionally favour the structural route to categorisation, according to which forming a description of a novel shape in terms of its parts and their spatial relationships is a prerequisite to the ability to categorise it. In comparison, we demonstrate that knowledge of instances of each of several representative categories can provide the necessary computational substrate for the categorisation of their new instances, as well as for representation and processing of radically novel shapes, not belonging to any of the familiar categories. The representational scheme underlying this approach, according to which objects are encoded by their similarities to entire reference shapes (S Edelman, 1997 Behavioral and Brain Sciences in press), is computationally viable, and is readily mapped onto the mechanisms of biological vision revealed by recent psychophysical and physiological studies.
