Shape from contours as initial step in shape from multiple cues
Thomas Fromherz, Martin Bichsel · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1994
This paper presents a method for converting the contours of a rotated object's profiles into a volumetric object description that is suitable for an integration of complementary visual shape cues. In contrast to traditional shape from contours algorithms, the method presented in this paper does not attempt to construct an `optimum' solution but rather finds an enclosing volume which must contain the true solution. In this way, an infinite set of solutions is compactly represented which is compatible with the contours in all available views. In the experimental set-up, the object is recorded from multiple viewing angles with a fixed camera. The recorded frames then pass a segmentation step which is carefully designed so that the true object volume will lie inside the reconstructed enclosing volume. The resulting segments are integrated into a voxel based description of the enclosing volume. For illustration purposes, the procedure is tested on a few general objects such as both a human and a dummy's head and a rigid plastic glove.