Published online by Cambridge University Press: 20 May 2010
Seeing is forgetting the name of the thing one sees.Paul Valéry (1871–1945)
If you are looking at the object, you need not think of it.Ludwig Wittgenstein (1889–1951)
A decisive resolution of the problems of high-level vision is at present impeded not by a shortage of computational ideas for processing the array of measurements with which vision begins, but rather by certain tacit assumptions behind the very formulation of these problems.
Consider the problem of object recognition. Intuitively, recognition means determining whether or not the input contains a manifestation of a known object, and perhaps identifying the object in question. This intuition serves well in certain contrived situations, such as character recognition in reading or machine part recognition in an industrial setting – tasks that are characterized first and foremost by only involving objects that come from closed, well-defined sets. An effective computational strategy for object recognition in such situations is to maintain a library of object templates and to match these to the input in a flexible and efficient manner (Basri and Ullman 1988; Edelman et al. 1990; Huttenlocher and Ullman 1987; Lowe 1987).
In categorization, in which the focus of the problem shifts from identifying concrete shapes to making sense of shape concepts, this strategy begins to unravel – not because flexible template matching as such cannot keep up with the demands of the task, but rather because the template library is no longer well-defined at the levels of abstraction on which the system must operate.