Search

A striking aspect of the human capacity for object recognition is that it is largely unaffected by where the object's image falls on the retina, the size of the image, or the orientation in depth from which the object is viewed. A fundamental challenge for theories of human object recognition is to explain how to achieve this invariance with viewpoint. In this article, the author reviews four general approaches to this problem in the computational modeling literature and discuss evidence bearing on each as an account of human object recognition. Some of the earliest computer models of object recognition were based on structural descriptions in 3-D object-centered reference frames. According to most early structural description theories, object recognition is accomplished by using the 2-D image of an object to derive a 3-D structural description, which is then matched against similar descriptions stored in memory. Some view-matching models exploit specific transformations, such as translation, scaling, and rotation, to bring novel images into correspondence with a small number of stored views.