Generalization to Novel Views: Universal, Class-based, andModel-based Processing

• Authors: Yael Moses;Shimon Ullman
• Venue: International Journal of Computer Vision
• Year: 1998

A major problem in object recognition is that a novel image of a given objectcan be different from all previously seen images. Images can vary considerablydue to changes in viewing conditions such as viewing position andillumination. In this paper we distinguish between three types of recognitionschemes by the level at which generalization to novel images takes place:universal, class, and model-based. The first is applicable equally to allobjects, the second to a class of objects, and the third uses known propertiesof individual objects. We derive theoretical limitations on each of the threegeneralization levels. For the universal level, previous results have shownthat no invariance can be obtained. Here we show that this limitation holdseven when the assumptions made on the objects and the recognition functions are relaxed. We also extend the results to changes ofillumination direction. For the class level, previous studies presentedspecific examples of classes of objects for which functions invariant toviewpoint exist. Here, we distinguish between classes that admit suchinvariance and classes that do not. We demonstrate that there is a tradeoffbetween the set of objects that can be discriminated by a given recognitionfunction and the set of images from which the recognition function canrecognize these objects. Furthermore, we demonstrate that although functionsthat are invariant to illumination direction do not exist at the universallevel, when the objects are restricted to belong to a given class, an invariantfunction to illumination direction can be defined. A general conclusion of thisstudy is that class-based processing, that has not been used extensively in thepast, is often advantageous for dealing with variations due to viewpoint andilluminant changes.