High-Quality Visualization and Filtering of Textures and Segmented Volume Data on Consumer Graphics Hardware
Author(s): Markus Hadwiger.
PhD Thesis: VRVis Research Center and Institute of Computer Graphics and Algorithms, Vienna University of Technology,
2004.
[BibTeX] 
Abstract:
Most rendering methods in visualization and computer graphics are focusing either on image quality in order to produce “correct” images with non-interactive rendering times, or sacrifice quality in order to attain interactive or even real-time performance. However, the current evolution of graphics hardware increasingly allows to combine the quality of off-line rendering approaches with highly interactive performance. In order to do so, new and customized algorithms have to be developed that take the specific structure of graphics hardware architectures into account. The central theme of this thesis is combining high rendering quality with real-time performance in the visualization of sampled volume data given on regular three-dimensional grids. More generally, a large part of this work is concerned with high-quality filtering of texture maps, regardless of their dimension. Harnessing the computational power of consumer graphics hardware available in off-the-shelf personal computers, algorithms that attain a level of quality previously only possible in off-line rendering are introduced. A fundamental operation in visualization and computer graphics is the reconstruction of a continuous function from a sampled representation via filtering. This thesis presents a method for using completely arbitrary convolution filters for high-quality reconstruction exploiting graphics hardware, focusing on real-time magnification of textures during rendering. High-quality filtering in combination with MIP-mapping is also illustrated in order to deal with texture minification. Since texturing is a very fundamental operation in computer graphics and visualization, the resulting quality improvements have a wide variety of applications, including static texture-mapped objects, animated textures, and texture-based volume rendering. The combination of high-quality filtering and all major approaches to hardwareaccelerated volume rendering is demonstrated. In the context of volume rendering, this thesis introduces a framework for high-quality rendering of segmented volume data, i.e., data with object membership information such as segmented medical data sets. High-quality shading with per-object optical properties such as rendering modes and transfer functions is made possible, while maintaining real-time performance. The presented method is able to filter the boundaries between different objects on-the-fly, which is non-trivial when more than two objects are present, but important for high-quality rendering. Finally, several approaches to high-quality non-photorealistic volume rendering are introduced, a concept that is especially powerful in combination with segmented volume data in order to focus a viewer’s attention and separate focus from context regions. High-quality renderings of isosurfaces are obtained from volumetric representations, utilizing the concept of deferred shading and deferred computation of high-quality differential implicit surface properties. These properties include the gradient, the Hessian matrix, and principal curvature magnitudes as well as directions. They allow high-quality shading and a variety of nonphotorealistic effects building on implicit surface curvature. We conclude that it is possible to bridge the gap between traditional high-quality offline rendering and real-time performance without necessarily sacrificing quality. In an area such as volume rendering that can be very demanding with respect to quality, e.g., in medical imaging, but whose usefulness increases significantly with higher interactivity, combining both high quality and high performance is especially important.