A SYSTEM FOR IMAGE/VIDEO OBJECT QUERY BY SHAPE - Video Object Extraction and Representation: Theory and Applications - page 128

segmentation work into its analysis, we treat video object extraction
and representation separately.
Some shape query methods treat the shape as a signal with no as-
sumptions about the content either as a 2-D area or as a 1-D contour.
Area-based methods such as moment-based shape descriptions [Prokop
and Reeves, 1992] [Khotanzad and Hong, 1990] [Bailey and Srinath,
1996] treat shape as the signal over a 2-D space while contour-based
method such as a Generalized Hough Transform [Jeng and Tsai, 1991]
[Samal and Edwards, 1997] and curvature-scale space [F. and A.K, 1992]
treat the shape as signal over 1-D order. However, even the choice of
an area-based vs. contour-based representation biases the concept of
similarity. For instance, thresholding has very different effects depen-
dent upon the shape representation. A long, but thin protrusion may be
removed by thresholding in an area-based representation while it may
be an important feature in a contour-based representation. While shape
seems to be an objective feature of a video object, its interpretation
depends strongly on the method of representation.
If we have some a priori knowledge about a class of objects, we can
leverage this knowledge into the design of our shape query system. If we
know how an object shape deforms, we can add invariance to these de-
formations into our shape query system. For instance, a template-based
shape matching allows full flexibility along the range of its parameters
[Fischler and Elschlager, 1973]. Our shape extraction is based upon pre-
vious concepts of Euclidean distance, shape thinning algorithms, and a
concept of Voronoi Skeletons [Cantoni and Carrioli, 1981] [Mayya and
Rajan, 1995] [Ogniewicz and Kubler, 1995]. The representation is a
novel recursive data structure based on work in Chapter 5 that allows
us to robustly represent these Voronoi Order Skeletons (VOS). Shape
query systems that use a priori knowledge can only be applied reliably
to a subset of object shapes: in our case, shapes that contain an physi-
cal skeletal structure. However, as long as we recognize the underlying
assumptions of our shape descriptor, we effectively apply this shape de-
scriptor to search for desired video content.
2. PROBLEM DESCRIPTION
Although the evaluation of system results for shape query is subjec-
tive, we provide a general shape query problem description, shape query
terms and guidelines on how to interpret results.
The problem of shape query is described as follows. In the context
of video object search of the MPEG-7 standard [Group, 1999], given
a query shape as a black and white image, find the shapes within the
database that are manually classified as the same class of content. The