The Evolution of Object Categorization and the Challenge of Image Abstraction
Sven Dickinson · Cambridge University Press eBooks · 2009
Introduction In 2004, I was a guest at the Center for Machine Perception at the Czech Technical University. During my visit, a graduate student was kind enough to show me around Prague, including a visit to the Museum of Modern and Contemporary Art (Veletrzní Palác). It was there that I saw the sculpture by Karel Nepras entitled “Great Dialogue,” a photograph of which appears in Figure 1.1. The instant I laid eyes on the sculpture, I recognized it as two humanoid figures seated and facing each other; when I've presented a 2-D image (Fig. 1.1) of the sculpture to classroom students and seminar audiences, their recognition of the two figures was equally fast. What's remarkable is that at the level of local features (whether local 2-D appearance or local 3-D structure), there's little, if any, resemblance to the features constituting real 3-D humans or their 2-D projections. Clearly, the local features, in terms of their specific appearance or configuration, are irrelevant, for individually they bear no causal relation to humans. Only when such local features are grouped, and then abstracted , do the salient parts and configuration begin to emerge, facilitating the recognition of a previously unseen exemplar object (in this case, a very distorted statue of a human) from a known category (humans).