Abstract

Our long-range goal is detecting instances from a large number of object classes in a computationally efficient manner. Detectors involving a hierarchy of tests based on edges have been used elsewhere and shown to be quite fast online. However, significant further gains in efficiency - in representation, error rates and computation - can be realized if the family of detectors is constructed from common parts. Our parts are flexible, extended edge configurations; they are learned, not pre-designed. In training, object classes are presented sequentially; the objective is then to accommodate new classes by maximally reusing parts. Ideally, the number of distinct parts in the system would grow much more slowly than linearly with the number of classes. Initial experiments on learning to detect several hundred T E Xsymbols are encouraging.

Citations