Local Data Models for Probabilistic Transfer Function Design

Abstract

A central component of expressive volume rendering is the identification of tissue or material types and their respective boundaries. To perform appropriate data classification, transfer functions can be defined in highdimensional histograms, removing restrictions of purely 1D scalar value classification. The presented work aims at alleviating the problems of interactive multi-dimensional transfer function design by coupling high-dimensional, probabilistic, data-centric segmentation with interaction in the natural 3D space of the volume. We fit variable Gaussian Mixture Models to user specified subsets of the data set, yielding a probabilistic data model of the identified material type and its sources. The resulting classification allows for efficient transfer function design and multi-material volume rendering as demonstrated in several benchmark data sets.