Disambiguation techniques for recognition in large databases and for under-constrained reconstruction

When computing 3D interpretations of noisy 2D images, many interpretations are often plausible. We describe a general framework for resolving such ambiguities in object recognition and reconstruction using maximum likelihood estimation. To this end we define measures of likelihood and stability of interpretations. These measures also give a practical way to evaluate how 'generic' views are, and identify 'characteristics' views. To demonstrate the usefulness and generality of this framework, we computed the proposed stability and likelihood measures using 4 different kinds of image matching algorithms, matching: (1) feature points, (2) angles, (3) occluding contours of smooth surfaces, and (4) shaded images of smooth surfaces.