Visual Exploration of Multivariate Volume Data Based on Clustering - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 173

(c)

(a)

(b)

Fig. 16.4 Coordinated views for visual exploration of clustering result: a Cluster tree visualization
of a hierarchical density-based clustering approach. b Parallel coordinates plots of selected clusters.
c 3D texture-based volume rendering of selected clusters with selected material properties. System
is applied to a physical ionization front instability simulation dataset with a ten-dimensional feature
space. (Data set provided in the 2008 IEEE Visualization Design Contest [ 41 ])

16.6 Interactive Modification of Clustering Result

User-guided clustering approaches like the ones by Ivanovska and Linsen [ 17 ]or
Tzeng and Ma [ 39 ] allow for a modification of the clustering result within their
user-guided clustering framework as described above. Since automatic clustering
approaches may also produce results that are subject to manual improvements and
since the user may bring in some domain expertise that may go beyond what can
be achieved with fully automatic approaches, it is also desirable to allow modifica-
tions of the automatic clustering results. In particular, it is desirable to couple the
visual exploration described in the previous section withmeans to interactively adjust
clusters. As the visual exploration process may lead to new insights about the clus-
ters under investigation, those new insights shall be documented, e.g., by splitting a
cluster into two smaller clusters.
The system of coordinated views by Dobrev et al. [ 10 ] as shown in Fig. 16.4 also
allows for an interactive modification of the clusters. First, since density values are
given, clusters can be shrunk (and grown again) by adjusting the density level that
corresponds to the density cluster. Second, the parallel coordinates plot also serves
as an interactive widget. Brushing on the individual axes of the parallel coordinates
plot induces a selection that is directly reflected in the linked volume renderer. If
the interactions on the parallel coordinates plot lead to the detection of a certain
subcluster within the selected cluster, the cluster can be split appropriately. Third,
clusters can be merged by selecting them in the cluster tree widget and applying a
merge operation.

16.7 Conclusions and Future Directions

We have presented approaches for multivariate volume visualization that are based
on clustering methods. Clustering is applied to capture the information in the multi-
dimensional attribute space of the multivariate volume data into regions of sim-