30
is exactly the optimal condition for the graph-cut to perform well. The two bone
data share this characteristics as the connectivity between bone fragments is low. As
indicated from Table 4.1, the time to segmenting each of the four data sets is less than
five minutes. The piggy bank and bunny data sets presented the most difficulty in our
tests. In all our tests, they were the largest in size. The time it took to perform single
min-cut segmentation was on the order of a minute, and this hindered interactivity
greatly. Fhrthermore, the piggy bank example was difficult due to the lack of distinct
separation between the coins the the inner surface of the pig. Also problematic was
the fact that the components were defined less by graph-connectivity and more by
curvature. Both of these factors Componded to slow the segmentation. The bunny
example shared the connectivity/curvature problem, which contributed as well to its
slower times. However, the two data sets were segmented in around fifteen minutes
even with the aforementioned difficulties.
4.2.2 Protein Data Bank Examples
We have segmented datasets from the Protein Data Bank (PDB) entries. The PDB
models are specified as a set of atomic coordinates, where each atom can have se-
mantics attached to them. From the atomic coordinates, we build a volume over the
dataset by convolving the atoms with a radial function, such as a gaussian [17]. For
our purpose, we simply invoked the EMAN function pdb2mrc [16] to construct a vol-
ume over the input PDB. For comparison, we use a molecular visualization program
called UCSF Chimera [21]. Chimera colors protein entries by chains. By examin-
ing the models visually, we reproduced the chain segmentation using our painting
interface. Figure 4.1 illustrates that we were able to reproduce the segmentation
closely.