Visual Exploration of Multivariate Volume Data Based on Clustering - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 172

exhibit the individual attribute values for the clusters, but also visually encodes how
homogeneous or heterogeneous a cluster is. It can even reveal whether a cluster splits
into subclusters in one of the attributes.
16.5 Coordinated Views for Visual Exploration
of Clustering Result
The visual encodings presented in the previous section all represent the clustering
result, but they do reveal different properties of the clusters: The object-space rep-
resentation exhibits the distribution of the clusters within the volumetric domain,
the cluster hierarchy encodes the structure of the clusters, a projection shows the
distribution of the clusters within the multi-dimensional attribute space, and the par-
allel coordinates allow for retrieving individual attribute values. Consequently, it is
desirable to have all those visual encodings embedded into a visual exploration sys-
tem. The system would provide different views on the data, where the views shall
be coordinated, i.e., any interactions like selection or filtering that are performed in
one view shall simultaneously also be applied to all the other views. Then, all views
provide one coherent snapshot of the data.
Akiba et al. [ 2 , 3 ] presented a system that operates with coordinated views on mul-
tivariate volume data. In particular, they used a representation of themulti-variate data
in parallel coordinates to allow the user to generate a transfer function by brushing
regions of interest. They even add another aspect to it, as they are dealing with time-
varying multivariate data, where another coordinated view shows a visual encoding
of changes over time. However, their approach is not based on the clustering idea
such that the amount of necessary user interaction increases with increasing dimen-
sionality and may at some point get cumbersome.
Dobrev et al. [ 10 ] follows the idea of using clustering to provide intuitive
operations in cluster space. Like for Akiba et al., the system is based on an intu-
itive user interface, but the combination with hierarchical density-based clustering
has the benefit that it scales better to high dimensionality. Object-space representa-
tion of the clusters is achieved using a GPU-based volume rendering approach, to
which only the cluster IDs and the density values from the clustering approach are
handed. Each cluster is assigned a unique color and opacity value according to the
user selections. The gradients of the density values can be used to obtain appropriate
surface normals for the clusters, which are needed for illumination. The cluster tree is
visually encoded as a radial nodelink diagram, which serves as the main interaction
widget for selecting clusters and assigning material properties. Parallel coordinates
plots are used to show the attribute values of the selected clusters. Figure 16.4 shows
the system with the three coordinated views.