The Evolution of Object Categorization and the Challenge of Image Abstraction

ion -at one extreme, the raw input image -at the other extreme, a categorical shape model -betwen them, a large representational gap ? -local appearance abstraction -local features map one-to-one to model features Figure 3: The evolution of object categorization over the past four decades (see text for discussion). directly (one-to-one) to the occluding boundaries (separating figure from background) and surface discontinuities of the high-order volumetric parts making up the models. The results left many unsatisfied, as the images and objects were often contrived (including blocks world scenes), and the resulting systems were unable to deal with real objects imaged under real conditions. Nevertheless, some very important principles emerged in the 1970’s, many of which are being rediscovered by today’s categorization community: 1. the importance of shape (e.g., contours) in defining object categories; 2. the importance of viewpoint-invariant, 3-D shape representations; 3. the importance of symmetry and other non-accidental relations in fea-