Faster Ray Tracing Using Adaptive Grids

fficient ray tracing has been a contradiction in terms since its introduction as a computer version of ideas found in Dürer's Underweysung der Messung (1525) and Descartes' La Dioptrique (1637). The most effective acceleration techniques developed to reduce ray tracing's high computational cost are based on space coherence: bounding box hierarchies and space subdivision. 1 During pre-processing, a space subdivision algorithm associates objects with the space elements in which they reside. During ray runtime, a ray traverses the volume elements and is tested for intersection only with the objects inhabiting those voxels. Partitioning 3D space into a regular mesh of voxels has attracted much attention. Fujimoto et al. 2 showed that in realistic scenes the method combined with an incremental ray traversal outpaces the adaptive octree approach by an order of magnitude. While octrees avoid excessive space subdivision in empty and sparsely populated regions, they require vertical traversal for detecting neighboring cells—a costly process even with improved methods, such as the Digital Differential Analyzer (DDA) octree traversal algorithm. The regular grid subdivision does not require vertical traversal, but incurs an unnecessary overhead of voxel-to-voxel steps in empty regions. To deal with the problem, Devillers 3 introduced macroregions, collections of axis-aligned boxes that replace empty voxel sets. Each empty voxel, possibly residing in several macroregions, points to the optimal macroregion, enabling the ray entering the voxel to leap over the largest number of empty voxels quickly. Yagel et al. 4 proposed a 3D raster ray tracer (RRT) that operates on a set of unit voxels associated with at most one object. To achieve a reasonable image quality, the voxel footprint has to be approximately pixel size. Even though now a ray-object intersection is found as soon as a ray enters a nonempty voxel, the method requires more than one billion voxels to achieve 1K resolution. This results in staggering memory requirements and an increased ray traversal cost. Cohen and Sheffer 5 mitigated the latter problem with proximity clouds—a method that stores in each subdivision cell the distance to the nearest nonempty cell— improving the RRT's performance by 30 percent in simple sparse scenes. The regular grid is conceptually simple, efficient, and easy to implement. However, it cannot meet the contradictory requirements of sufficient space decomposition in the densely populated areas and the need to avoid excessive partitioning in a scene's void regions. Although alternative solutions that mix space subdivision types accelerate …