Shape from contour by generating synthetic texture patterns

Shape from X techniques in machine vision use different cues for extracting depth information such as shape from contour, texture orientation, shading, stereo images, and motion. Contours are the most basic representation of object shape since it contains strong, dominant characteristics for determining object shape. This research deals with the problem of computing the 3-D surface from a single contour image of 3-D space curves made from thin sheets. Our approach uses the geodesic contour as the major cue for surface recovery, with the basic assumption that if a planar sheet is distorted, its imaged contour will represent the configuration of the underlying surface. There are three primary processes in our approach. First, the geodesic contour is used to approximate a set of intersection grid points via a relaxation procedure using the Gibbs Sampler with Simulated Annealing. Next, a spline technique forms a texture pattern accross the contour by using the grid of intersecting points obtained from the previous step as features. Finally, we use the classical shape from texture orientation technique to compute the surface-orientation at each grid position. Thus, we report a dense map of surface normals at various points within the contour to indicate the computed shape. Our approach can be used for both parallel and non-parallel geodesic contours. Our experimental results show that this method is effective in computing shape from contour of developable surfaces.