Towards Automatic Feature-based Visualization

Abstract

Visualizations are well suited to communicate large amounts of complex data. With increasing resolution in the spatial and temporal domain simple imaging techniques meet their limits, as it is quite difficult to display multiple variables in 3D or analyze long video sequences. Feature detection techniques reduce the data-set to the essential structures and allow for a highly abstracted representation of the data. However, current feature detection algorithms commonly rely on a detailed description of each individual feature. In this paper, we present a feature-based visualization technique that is solely based on the data. Using concepts from computational mechanics and information theory, a measure, local statistical complexity, is defined that extracts distinctive structures in the data-set. Local statistical complexity assigns each position in the (multivariate) data-set a scalar value indicating regions with extraordinary behavior. Local structures with high local statistical complexity form the features of the data-set. Volume-rendering and iso-surfacing are used to visualize the automatically extracted features of the data-set. To illustrate the ability of the technique, we use examples from diffusion, and flow simulations in two and three dimensions.