Certified by................ J-

The problem of comparing and aligning three-dimensional models to images has at-tracted a significant amount of interest in the field of computer vision. However, it has turned out to be a complex and difficult task. Traditional approaches to align-ment are plagued by a variety of pitfalls, which mainly stem from the complex nature of the imaging process. The image of an object is a function of many parameters, including the properties and position of the object as well as extraneous factors, like illumination conditions. The variations in illumination, in particular, have a strong effect on the image of an object but cannot be easily incorporated in an aligning algorithm. A recent alignment method, however, tackles the problem with greater success using a radically different approach. This method offers a more appropriate measure for comparing such dissimilar entities as image intensities and object properties, based on information theory. Alignment is achieved by maximizing an estimate of the mutual information between image and model using a stochastic gradient procedure