Pattern Search for the Visualization of Scalar, Vector, and Line Fields
The main topic of this thesis is pattern search in data sets for the purpose of visual data analysis. By giving a reference pattern, pattern search aims to discover similar occurrences in a data set with invariance to translation, rotation and scaling. To address this problem, we developed algorithms dealing with different types of data: scalar fields, vector fields, and line fields. For scalar fields, we use the SIFT algorithm (Scale-Invariant Feature Transform) to find a sparse sampling of prominent features in the data with invariance to translation, rotation, and scaling. Then, the user can define a pattern as a set of SIFT features by e.g. brushing a region of interest. Finally, we locate and rank matching patterns in the entire data set. Due to the sparsity and accuracy of SIFT features, we achieve fast and memory-saving pattern query in large scale scalar fields. For vector fields, we propose a hashing strategy in scale space to accelerate the convolution-based pattern query. We encode the local flow behavior in scale space using a sequence of hierarchical base descriptors, which are pre-computed and hashed into a number of hash tables. This ensures a fast fetching of similar occurrences in the flow and requires only a constant number of table lookups. For line fields, we present a stream line segmentation algorithm to split long stream lines into globally-consistent segments, which provides similar segmentations for similar flow structures. It gives the benefit of isolating a pattern from long and dense stream lines, so that our patterns can be defined sparsely and have a significant extent, i.e., they are integration-based and not local. This allows for a greater flexibility in defining features of interest. For user-defined patterns of curve segments, our algorithm finds similar ones that are invariant to similarity transformations. Additionally, we present a method for shape recovery from multiple views. This semi-automatic method fits a template mesh to high-resolution normal data. In contrast to existing 3D reconstruction approaches, we accelerate the data acquisition time by omitting the structured light scanning step of obtaining low frequency 3D information.