ABSTRACT
How do the concepts of pattern theory relate to existing computer vision theories of recognition? As we have seen, the image of an object depends on many imaging factors such as lighting conditions, viewpoint, geometric deformations of the object, albedo of the object, and whether it is partially occluded by other objects. In fact, one very careful study (Moses, Adini and Ullman 94) compared, using a number of different measures, the degree of variability of images of male faces (without glasses or beards and excluding hair) due to the three factors: (a) illumination, (b) viewpoint, and (c) different individuals. Their conclusion was that individual differences caused the least amount of change, viewpoint changes some 20% more, and illumination changes a whopping 150% more variability in the face image. Clearly, to recognize a particular face you must accomplish this in spite of a huge amount of variation from other factors. Existing theories of object recognition differ on how they deal with these factors. Do they follow the "analysis by synthesis approach" and invert the image formation process to determine what object(s) are most likely to have generated the image? How much use do they make of prior, or class specific, knowledge about particular objects? What types of transformations do they allow on the images?
