Generic Object Recognizer Design

The problem of generic object recognition is discussed in relation to the current research emphasis in the computer vision community on learning methods for classz~cationM. utual information is proposed as a tool for identiying the salient features of a class and as a mechanism for constructing class recognizers from an engineering design standpoint. The concept of obsewability is introduced to define classes that may differ from human concepts but are necessary to achieve high recognition performance.