Structure and Motion from Images of Smooth Textureless Objects

This paper addresses the problem of estimating the 3D shape of a smooth textureless solid from multiple images acquired under orthographic projection from unknown and unconstrained viewpoints. In this setting, the only reliable image features are the object’s silhouettes, and the only true stereo correspondence between pairs of silhouettes are the frontier points where two viewing rays intersect in the tangent plane of the surface. An algorithm for identifying geometrically-consistent frontier points candidates while estimating the cameras’ projection matrices is presented. This algorithm uses the signature representation of the dual of image silhouettes to identify promising correspondences, and it exploits the redundancy of multiple epipolar geometries to retain the consistent ones. The visual hull of the observed solid is finally reconstructed from the recovered viewpoints. The proposed approach has been implemented, and experiments with six real image sequences are presented, including a comparison between ground-truth and recovered camera configurations, and sample visual hulls computed by the algorithm.