Abstract

We claim that the task of object recognition necessitates a measure for image likelihood, that is: a measure for the probability that a given image is obtained from a familiar (pre-investigated) object. Moreover, in a system where objects are represented by 2D images, the best performance is achieved if those images are selected according to a maximum likelihood principle. This is equivalent to maximum stability of the image, or minimal change under a viewpoint perturbation. All of those qualities involve a quantitative comparison of similarity between images. We propose different metric functions which can be imposed on the image space of curved three dimensional objects. We use these metrics to detect the representative views (most stable and most likely views) of three test models. We find the same representative views under all the investigated metrics, suggesting that local maxima of stability and likelihood are metric independent. Our method of image comparison is based solely ...

Citations