Real-Time High-Resolution Sparse Voxelization with Application to Image-Based Modeling

Abstract

We present a system for real-time, high-resolution, sparse voxelization of an image-based surface model. Our approach consists of a coarse-to-fine voxel representation and a collection of parallel processing steps. Voxels are stored as a list of unsigned integer triples. An oracle kernel decides, for each voxel in parallel, whether to keep or cull its voxel from the list based on an image consistency criterion of its projection across cameras. After a prefix sum scan, kept voxels are subdivided and the process repeats until projected voxels are pixel size. These voxels are drawn to a render target and shaded as a weighted combination of their projections into a set of calibrated RGB images. We apply this technique to the problem of smooth visual hull reconstruction of human subjects based on a set of live image streams. We demonstrate that human upper body shapes can be reconstructed to giga voxel resolution at greater than 30 fps on modern graphics hardware.