A new shape representation for fast 3D reconstruction from multiple 2D perspective images

The reconstruction of 3D shape from 2D visual images is a primary goal of computer vision. Active methods such as range finding or laser striping are accurate but require expensive equipment. This has motivated work in the area of passive techniques which seek to infer 3D depth information from one or more 2D intensity images. Work in this area includes shape from shading, texture, stereo, contours and motion. Of these none provided the accuracy of active methods, however shape from multiple contour methods such as silhouette intersection are able to reconstruct good approximations to object shape when a good segmentation of the scene object can be made. In the silhouette intersection process each contour is used to construct a back projection cone. The cone is formed by rays passing from the optical centre of the camera through the contour in the image plane. The object is constrained to lie within this cone. When a number of images observed from known positions are available the intersection of the set of defined cones describes a volume which encompasses the object. As the number of views is increased this volume approaches what is called the object´s visual hull. For many objects this is congruent to the objects true shape. For many others it is a good approximation to the true shape. This paper is concerned with methods to represent the object silhouettes and back projection cones