Object Categorization: The Evolution of Object Categorization and the Challenge of Image Abstraction

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