5 : Experiment
In this section, we provide some experiment results for our methods. Our
process can be simply described as below:
• For a given binary shape, we first squeeze and enlarge it to construct the
hierarchy scale levels from small to big.
Fig. 4 Sample views of the silhouette objects
Fig. 5 Example for our methods
• Perform morphological shape decomposition for each level, in this paper
we use Pitas and Venetsanopoulos [14] method. Find the corresponding
matching parts through all levels. These parts input for the next step.
• Use the output from last step to construct the graph structure. Use average
graph energy method to delete the noise nodes(parts) in the graph. Repeat
this step until the average graph energy never rise. Output the final graph
structure.
We experiment on shock graph database which composed of 150 silhouettes
of 10 kinds of objects [16]. An example of database is shown in Figure 4.
In Figure 5, we give some results for our methods, here in the left column
is the origin shape, the middle column is the pruned skeleton parts from morphological
shape decomposition and the right column is the re-constructed
Shape Decomposition for Graph Representation 9
Table 1 Variation for the number of parts with different shape decomposition
methods.
Class Name MSD Our Method
Car 8.5 4.1
Children 11.4 6.7
Key 9.0 5.0
Bone 8.5 4.7
Hammer 4.5 3.2
shape by using the skeleton centers in the middle column. From this example,
we can see that our algorithm can reduce some noise parts from the origin
morphological decomposition while keep the important parts. It can be seen
that the reconstructed shapes are quite similar to the original shapes and
thus keeps the most important information for further analysis.
In table 1 we listed the variation for the number of parts within the same
class for tradition morphological shape decomposition method(MSD) and our
method. It is clear that the variations for the number of parts for the tradition
morphological shape decomposition is higher than our method.
In this section, we provide some experiment results for our methods. Our
process can be simply described as below:
• For a given binary shape, we first squeeze and enlarge it to construct the
hierarchy scale levels from small to big.
Fig. 4 Sample views of the silhouette objects
Fig. 5 Example for our methods
• Perform morphological shape decomposition for each level, in this paper
we use Pitas and Venetsanopoulos [14] method. Find the corresponding
matching parts through all levels. These parts input for the next step.
• Use the output from last step to construct the graph structure. Use average
graph energy method to delete the noise nodes(parts) in the graph. Repeat
this step until the average graph energy never rise. Output the final graph
structure.
We experiment on shock graph database which composed of 150 silhouettes
of 10 kinds of objects [16]. An example of database is shown in Figure 4.
In Figure 5, we give some results for our methods, here in the left column
is the origin shape, the middle column is the pruned skeleton parts from morphological
shape decomposition and the right column is the re-constructed
Shape Decomposition for Graph Representation 9
Table 1 Variation for the number of parts with different shape decomposition
methods.
Class Name MSD Our Method
Car 8.5 4.1
Children 11.4 6.7
Key 9.0 5.0
Bone 8.5 4.7
Hammer 4.5 3.2
shape by using the skeleton centers in the middle column. From this example,
we can see that our algorithm can reduce some noise parts from the origin
morphological decomposition while keep the important parts. It can be seen
that the reconstructed shapes are quite similar to the original shapes and
thus keeps the most important information for further analysis.
In table 1 we listed the variation for the number of parts within the same
class for tradition morphological shape decomposition method(MSD) and our
method. It is clear that the variations for the number of parts for the tradition
morphological shape decomposition is higher than our method.
No comments:
Post a Comment