RastaVox: Memory-Efficient Voxel-Model Rasterization

Voxel models are of increasing interest in 3D computer games, as they give rise to many efficient operations that can not be performed easily on adaptive surface models like triangle meshes. Examples include the efficient generation of level of detail hierarchies and model modifications like carving and extruding. However, to faithfully represent the surface and, thus, to render it in a similar quality than polygonal meshes, voxel representations require high spatial resolution along the surface. This typically implies that voxel models take up huge amounts of memory, making them unsuitable for computer games which have to run on fixed memory budgets. In this report, we propose a novel approach for constructing, representing, and rendering voxel models efficiently on recent GPUs, at memory budgets on par with the requirements in computer games. We achieve this by adapting and extending classical iso-surface rendering techniques to modern GPUs, including the on-the-fly reconstruction of surface-aligned voxel models from volume data. Unlike previous techniques our approach utilizes the GPU’s rasterization units for rendering. This makes it easy to integrate into existing rendering pipelines and allows taking advantage of hardware accelerated anti-aliasing. We demonstrate the construction of level of detail hierarchies for high quality rendering and real-time editing operations directly on our compact voxel representation.