Exact structure from motion is an ill-posed computation and therefore is sensitive to noise. It is shown here that a qualitative shape feature, the classification of surfaces into convex, concave, cylindrical, planar, and hyperbolic regions, can be computed directly from motion disparities, without the computation of an exact depth map. It is also shown that humans can judge the curvature sense of three points undergoing apparent motion from two, three, and four views with success rate significantly above chance.
Bibliographical noteFunding Information:
I am very grateful to E. Hildreth and R. Basri for many useful com-mentsr egardingt he manuscript,a nd to S. Edelmanw ho helpedm e with the implementationa nd analysis of the psychophysicale xperiments.I also thank H. Biilthoff, N. Cornelius, M. Dornay, M. Fahle, S. Kirkpa-trick, M. Ross, and A. Shashua.T his researchw as done partly in the MIT AI Laboratory. It was supportedb y a Fairchild postdoctoralf el-lowship, in part by grants from the office of Naval Research (N00014-88-k-0164a) nd the National Science Foundation( IRI-8719394a nd IRI-8657824)a nd by a gift from the James S. McDonnell Foundation to Professor E. Hildreth.