Resumen de Matching fabric appearance by primitives
Realism in fabric rendering depends on light scattering at fiber level and very detailed representation models. Nowadays, three representation models exist: based on surfaces, volumetric data and explicit geometry. The most accurate are both volumetric and explicit geometry at fiber scale, because they keep the appearance independently of view distance. However, due to their demanding hardware requirements and inherent algorithmic complexity, only a small group of fabrics are representable virtually. Therefore, we focus on developing new models by increasing their expressiveness and modeling features.
First, we focus on making a deep state-of-the-at about fabric rendering. Fabric rendering is growing in importance, motivated by textile and entertaiment industry. In this section, we summarize and evaluate the latest techniques involved in fabric rendering like fiber scattering, representation models, and parameter acquisition. Moreover, we discuss the underlying physical structures that compose a fabric.
Second, we intoduce a novel fabric appearance model taking into account fabrication properties like hue or fiber shape. This model is the first approach for predictive rendering in fabrics showing the lack of representation of previous approaches. Later, we introduce two novel explicit geometry representation models. Firstly, we continue the approach of previous works relying on generative function with pseudo-random noises. However, these models are too limited due to their lack of representation and editability. As a consequence, we have developed a second algorithm based on rules inspired by procedural generative models. This model is able to create novel complex shapes for fibers and yarns.
Finally, this work introduces improvements to the latest GPU voxelizer methods by raster. We analyze and propose solutions to their most important issues. Because of their inherent limitations, we introduce a novel GPU voxelization algorithm in Cuda. Our approach is specialized in fibrous materials with hundreds of millions of samples by model. With this algorithm, we can generate dense histograms, while maintaining the spatial information and optimizing the memory footprint. This voxelizer opens the door to new research lines by studying the fiber distributions in different yarns and weaving patterns.
