Primitives, Perceptual Organization and Object Recognition.

David Forsyth, Jitendra Malik, Margaret Fleck and Jean Ponce.

Submitted to Vision Research (special issue on Models of Recognition), February 1997.


We argue that any computational theory of object recognition should explain how to decide what is to be recognized and how to recognize objects at the level of abstract categories. We show that several current theories perform poorly by these criteria and propose that object representation must be in terms of the spatial relationships between stylized primitives, while the fundamental process of recognition is to assemble these primitives from image evidence, a version of perceptual organization. We illustrate our approach with programs aimed at finding people and horses in pictures, and discuss the role of learning in the corresponding recognition process.