Digital Signal Processing Reference
In-Depth Information
A System for Image
/
Video Object Query By Shape
of a dog shape finds other dogs, but mostly other animals that have
four legs. The second column shows the flexibility of the shape query
algorithm, allowing the shape to deform while still maintaining a ro-
bust similarity measure. The third column shows the overreliance of the
similarity criterion on the implicit support structure of the shape. To
match the protrusions of the two-humped camel of the third column, our
query system finds other shapes with just as many protrusions, rather
than one-humped camels. We can see that structure takes priority over
the actual shape of the animal. It is interesting to note that only two-
humped camels score very well, though. The fourth column shows results
of a shape that does not follow our implicit structure assumptions. As
expected, the query brings rather noisy results.
For the objective measures of MPEG-7 core experiments, there were
three different parts of experiment designed to test different qualities
of the shape description system. Part A tested the rotation and scale
invariance of the shape descriptor. Database A is comprised of 60 dif-
ferent images, each from a different class. To test rotational invariance,
each image is copied into the database at five different rotations of 9,
36, 45, 90 and 150 degrees; the original image is then used to search the
database for these 5 rotations along with the original image. A query
with the original image is done and the top six choices are matched to
whether they were derived from the original image. The percentage of
this matching is returned. A similar test is done for scale invariance,
adding scaled images of the database images at scales of 0.1, 0.2, 0.25,
0.3 and 2. The result of Test A is the average of scale and rotational
invariance tests. Test B tests the classification abilities of the shape de-
scriptor. Test B has a database of 60 different classes with 20 images
per class. For every image of each class, the top 40 images are found.
The final score of Test B is the average percentage of images that are
from the same class that are found in the top 40. Part C tested the
tolerance to video motion of the shape descriptor. Outlines from a 200
consecutive frames of video sequence of a Bream fish is mixed with 1100
outlines of other fish. The final score of Test C is the percentage of the
top 200 images in the query that are of the Bream fish.
Our comparative results are given in Figure 6.17. While our results are
not the best, we feel that the method is competitive with current state
of the art techniques that were presented in MPEG-7 Core Experiment
on Shape.
7. CONCLUSION
The MPEG-7 standard wishes to form a representational language for
video objects to support the foundation of video object search engines.
Search WWH ::
Custom Search