29
|
model |
volume size |
nodes |
segs |
min-cut |
total |
|
engine |
256 × 256 × 256 |
54485 |
7 |
8.89 |
123 |
|
foot |
256 × 256 × 256 |
337546 |
19 |
27.22 |
157 |
|
head |
128 × 256 x 256 |
368512 |
7 |
7.535 |
234 |
|
GroEL |
240 × 240 × 240 |
589863 |
14 |
23.83 |
200 |
|
Hsp26 |
128 × 128 × 128 |
110875 |
13 |
3.05 |
95 |
|
pɪg |
512 × 512 × 134 |
5109185 |
6 |
69.93 |
888 |
|
bunny |
512 × 512 x 361 |
1825425 |
7 |
73.37 |
971 |
Table 4.1 : The timing results for segmenting various data sets. The “segs” column
represents the number segments in the output. The “nodes” column represents the
number of nodes in the induced graph. The min-cut column is the time (in seconds)
to perform min-cut on the graph. The “total” column represents the total time (in
seconds) to perform the segmentation.
4.2.1 Heterogenous Examples
Table 4.1 shows the approximate times it took to segment the different data sets. We
chose 7 test sets of varying sizes and domains. We consider the user experience time
in painting these data sets. Each model is painted three times with the best time
being recorded. The painting time includes all processing time and the segmentation
time. The processing time depends largely on the size of the segmentation graph and
the number of segments. In our largest test case, the bunny, the induced graph has
over one milion nodes and seven segments. The graph-cut algorithm for the bunny
model runs on the order of a minute.
Except in the case of the two molecular structures, we do not have any expert
knowledge on the proper segmentation; in most cases, we do not think there is an
absolute baseline comparison. Therefore, we segment out components that we feel
is an intuitive component of the whole. In the cases of molecular data, we have
prior knowledge of the separation of the GroEL and Hsp26 structures into symmetric
subunits.
The molecular data sets are the least difficult to segment. They have the property
that corresponding graph connectivity between each symmetric subunit is low; this