Interactive Multiresolution and Multiscale Visualization of Large Volume Data
Interactive visualization and analysis of large and complex volume data is an ongoingchallenge. Data acquisition tools produce hundreds of Gigabytes of data andare one step ahead of visualization and analysis tools. Therefore, the amount ofdata to be rendered is typically beyond the limits of current computer and graphicshardware performance. We tackle this challenge in the context of state-of-the-artout-of-core multiresolution volume rendering systems by using a common mathematicalframework (a) to extract relevant features from these large datasets, (b) toreduce and compress the actual amount of data, and (c) to directly render/visualizethe data from the framework coefficients.This thesis includes an extended state-of-the-art analysis of data approximationapproaches and how they can be applied to interactive volume visualizationand used for feature extraction. Data is often approximated or reduced by usingcompact data representations, which require fewer coefficients than the originaldataset. In this thesis, the higher-order extension of the matrix singular value decompositionsummarized under the term tensor approximation (TA) was chosenas compact data representation. Tensor approximation consists of two parts: (1)tensor decomposition, usually an offline process, to compute the bases and coefficients,and (2) tensor reconstruction, typically a fast real-time process that invertsthe decomposition back to the original data during visualization.From these basic concepts, we derive how multiresolution volume visualizationand multiscale feature extraction are linked to the tensor approximationframework. The two axes of the TA bases were chosen as handles for multiresoluiiiivtion and multiscale visualization. The properties along the vertical axis of the TAbases match well the needs of state-of-the-art out-of-core multiresolution volumevisualization, where different levels of detail are represented by coarser or higherresolution representations of the same dataset and portions of the original datasetare loaded on demand in the desired resolution. Thus, the vertical axis of the TAbases is used for spatial selectivity and subsampling of data blocks. The horizontalaxis of the TA bases makes it possible to reconstruct the dataset at multiplefeature scales through the so-called tensor rank. Choosing only a few ranks correspondsto a low-rank approximation (many details removed) and choosing manyranks corresponds to an approximation more closely matching the original. Furthermore,a feature scale metric was developed to automatically select a featurescale and a resolution for the final reconstruction. In this scenario, the user selectsa desired feature scale for the approximated data, which is then used by the visualizationsystem to automatically define the resolution and the feature scale forthe current view on the dataset.Thanks to the compact data representation by TA, a significant data compression(15 percent of the original data elements) was achieved, which keeps the storagecosts low and boosts the interactive visualization. The interactive visualizationis moreover accelerated by using GPU-based tensor reconstruction approaches.The viability of interactive multiscale and multiresolution volume visualizationis tested with different TA volume visualization frameworks: (1) a simplebricked TA multiresolution, and (2) a TA multiresolution framework that usesglobal tensor bases. Both TA frameworks build on the vmmlib tensor classes,which were specifically developed for this thesis. For the testing, large volumedatasets from micro-computed tomography (microCT) and phase-contrast synchrotrontomography (pcST) that range up to 34 Gigabytes were acquired. Weshow visual as well as computational comparisons to state-of-the-art approachessuch as wavelet transform.We conclude by pointing out the tensor approximation framework to be a unifiedframework for interactive multiscale and multiresolution volume visualizationsystems, which directly controls data approximation in terms of feature scale andmultiple levels of detail. To wrap up, we discuss the achieved results and outlinepossible future work directions.