A two-level clustering approach for multidimensional transfer function specification in volume visualization

Abstract

Multidimensional transfer functions can perform more sophisticated classification of volumetric objects compared to 1-D transfer functions. However, visualizing and manipulating the transfer function space is non-intuitive when its dimension goes beyond 3-D, thus making user interaction difficult. In this paper, we propose to address the multidimensional transfer function design problem by taking a two-level clustering approach, where the first-level clustering by the self-organizing map (SOM) projects high-dimensional feature data to a 2-D topology preserving map, and the second-level clustering on the SOM neurons reduces the design freedom from a large number of SOM neurons to a manageable number of clusters. Based on the two-level clustering results, we propose a novel volume exploration scheme that provides top-down navigation to users exploring the volume. Guided by an informative volume overview, interesting structures in the volume are discovered interactively by the user selecting clusters to visualize and modifying the clustering results when necessary. Our interface keeps track of each interesting structure discovered, which not only enables users to inspect individual structures closely, but also allows them to compose the final visualization by fusing the structures deemed important.