Time-Critical Visual Exploration of Scalably Large Data

Abstract

This paper discusses visualization and analysis issues as datasets grow towards very large sizes, and it develops an approach to attack them. Datasets of this size become exploration-dominant since the scientists who create or collect them do not know, in detail, what's inside. Thus the methods developed here support exploratory visualization. To be fully successful these methods must be fast, so issues of time-criticality are addressed. Fast global overviews are prepared automatically based on an analysis of patterns in the data. From these particular overviews can be generated followed by detailed subviews, where these last steps are controlled by user interaction. A particular approach is developed to recognize spatial clustering in 3D data, and this is applied to a variety of datasets. The performance of the approach as a function of dataset size is analyzed, and it is found that it holds promise for the exploration of large datasets. In addition an octree decomposition method is also developed as an adjunct to the clustering method. Both methods can be used to develop hierarchical structures for the datasets that can be extended by user interaction. Information derived from the methods can be analyzed so that patterns in the datasets can be segmented according to shape, size, dynamic behavior, or content.