Abstract

The problem of recognizing complex articulated objects seems to require nonlocal nonhomogeneous computation. As a result, it is more difficult to imagine a neural net architecture that performs such a task. To gain insight into this problem we developed a prototype system for recognizing relatively complex articulated objects from their projected 2-D silhouettes by use of a massively parallel network of relatively simple processors. The computation is accomplished by breaking the recognition process down into a series of progressively more abstract estimation problems, each of which can be solved by local interaction between relatively simple computational units.

© 1988 Optical Society of America