Enforcing Scene Constraints in Single View Reconstruction

Abstract

Three-dimensional reconstruction from a single view is an under-constrained process that relies critically upon the availability of prior knowledge about the imaged scene. This knowledge is assumed to be supplied by a user in the form of geometric constraints such as coplanarity, parallelism, perpendicularity, etc, based on his/her interpretation of the scene. In the presence of noise, however, most of the existing methods yield reconstructions that only approximately satisfy the supplied geometric constraints. This paper proposes a novel single view reconstruction method that provides reconstructions which exactly satisfy all user-supplied constraints. This is achieved by first obtaining a preliminary reconstruction and then refining it in an extendable, constrained optimization framework.