Generalized voxel coloring

Abstract

Image-based reconstruction from randomly scattered views is a challenging problem. We present a new algorithm that extends Seitz and Dyer’s Voxel Coloring algorithm. Unlike their algorithm, ours can use images from arbitrary camera locations. The key problem in this class of algorithms is that of identifying the images from which a voxel is visible. Unlike Kutulakos and Seitz’s Space Carving technique, our algorithm solves this problem exactly and the resulting reconstructions yield better results in our application, which is synthesizing new views. One variation of our algorithm minimizes color consistency comparisons; another uses less memory and can be accelerated with graphics hardware. We present efficiency measurements and, for comparison, we present images synthesized using our algorithm and Space Carving.