Abstract
Given a 3D object I find a part decomposition accounting for observed perceptual similarities among the object's views. This is the inverse, ill-posed version of the direct problem of computing view similarity when the object parts are known. The problem is solved by inverting a proposed model for the direct similarity-from-parts problem. The algorithm takes as input the geometry of the object, the camera orientations corresponding to the test views, and the perceptual similarities among the views collected in psychophysical experiments. The goal is to find a part decomposition compatible with the observed view similarities. This work extends the results reported in two previous papers (Neural Computation 11, CVPR 2000). The major contribution reported here is the introduction of a 2D parameterization of the surface of the 3D objects. This allowed the use of an explicit, low-dimensional model for the shape of the surface regions defining the object parts. Specifically, each object part was identified with a Gaussian defined on the object's surface. The parameters of the surface Gaussians were adjusted to fit the measured view similarities, thus specifying a segmentation of the 3D object surface object into connected ellipsoidal regions, i.e. parts. The performance of the algorithm will be illustrated with experimental results.