EXISTING TECHNIQUES OF VISUAL PROCESSING - Video Object Extraction and Representation: Theory and Applications - page 43

CLASSIFICATION
When using a priori knowledge of an object for representation or ex-
traction, we require classification technologies. If we choose the correct
object representation, then a priori knowledge from an object represen-
tation can assemble partial information derived from the video sequence
with better noise rejection and tolerance to missing visual information. If
the object class is not known, incorrect application of a representation
actually degrades analysis by disallowing otherwise valid video object
features or analysis. For instance, the fitting of a human walker model
to the video sequence can locate all body parts of the walker and even his
gait, but only when the person is walking perpendicular to the camera
[Cheng and Moura, 1999]. How to determine whether the applicability
of a model and the reliability of results are still an open question. In
contrast, temporal smoothness applies to most objects and, if it is not
completely correct, degrades system performance in a graceful manner.
The issue of object classification can be avoided either 1) by the gen-
erality of the model or 2) by restricting the input to a certain object
class. These two options either limit the expressiveness of the model by
enforcing generality as a constraint on the model or fail to balance the
specificity of a model over a wide range of data, respectively. Simple
generalizations about the projected object behavior may mislead in a
small percentage of cases, but provide essential error resilience and re-
jection in the rest. By restricting the input class, the presupposition of
object class only reinforces the importance of a working classifier.
Classification technologies are not only important tool for extraction
and recognition, but also an important query functionality for the “new”
media of the future. In extraction, classification of unknown objects
allows the selection of the proper object-specific model that, in turn,
leverages object-specific a priori knowledge into our analysis. In rep-
resentation, such a classification technique can also be used as a query
tool. In Chapter 3, we present the concept of Voronoi Ordered Spaces
that fulfills the dual purpose of extraction and representation for both
our video object extraction and query.
5. DYNAMIC PROGRAMMING
The dynamic programming technique for optimization is used both in
Chapter 4 for VOP extraction and Chapter 6 for shape query. In 1955,
R. Bellman introduced the concept of dynamic programming for opti-
mization of certain problems [Bellman and Dreyfus, 1962] [Amini et al.,
1990]. It is a powerful and computationally efficient method for opti-
mization that can not only find the optimal solution for the optimization