ABSTRACT

Core problems in high-level vision

Shimon Edelman, University of Sussex

• Recognition: how to deal with novel views of shapes.
• Categorization: how to deal with novel instances of shape categories.
• Representation: how to deal with radically novel inputs which may belong to none of the familiar categories.
• Reasoning about structure: how to encode and refer to
  • the arrangement of parts in an object;
  • the arrangement of objects in a scene.

Theories of high-level vision normally hold that at least some of the aspects of these problems should be addressed by using symbols: auxiliary stand-in entities that

1. refer to events in the visual world without resembling them in any way, and
2. enter into discrete structural relationships with each other.

I shall offer a unified framework, based on the modeling of smooth low-dimensional view- and shape-manifolds, that may eventually provide an alternative to the structural approach, for each of the problems mentioned above.

Computer Science Home Page