On sparse voxel DAGs and memory efficient compression of surface attributes for real-time scenarios

Doctoral thesis, 2022

The general shape of a 3D object can expeditiously be represented as, e.g., triangles or voxels, while smaller-scale features usually are parameterized over the surface of the object. Such features include, but are not limited to, color details, small-scale surface-normal variations, or even view-dependent properties required for the light-surface interactions. This thesis is a collection of four papers that focus on new ways to compress and efficiently utilize surface data in 3D for real-time usage.
In Paper IA and IB, we extend upon the concept of sparse voxel DAGs, a real-time compression format of a voxel-grid, to allow an attribute mapping with a negligible impact on the size. The main contribution, however, is a novel real-time compression format of the mapped colors over such sparse voxel surfaces.
Paper II aims to utilize the results of the previous papers to achieve uv-free texturing of surfaces, such as triangle meshes, with optimized run-time minification as well as magnification filtering.
Paper III extends upon previous compact representations of view dependent radiance using spherical gaussians (SG). By using a convolutional neural network, we are able to compress the light-field by finding SGs with free directions, amplitudes and sharpnesses, whereas previous methods were limited to only free amplitudes in similar scenarios.