A SYSTEM FOR VIDEO OBJECT SEGMENTATION - Video Object Extraction and Representation: Theory and Applications - page 73

Figure 4.3. Surface-Representation of our video object: three different perspectives
on the surface-based representation of video object of the large coastguard ship from
the coastguard sequence frames 90-99.

4. The objects do not have a high degree of occlusion. Since we

do not model interactions between objects, our analysis cannot handle
a high degree of occlusion, i.e., when one object covers another from
view. In the future, inter-object relationships will be an important
aspect of the visual processing problems for both description and
extract ion.

3. PROBLEM FORMULATION VIA
SURFACE OPTIMIZATION
The surface optimization framework provides a convenient mathemat-
ical formulation that can integrate many sources of information and is
not bound to a particular computing architecture. We first treat the case
of a single video object in isolation. We decompose the multiple video
object segmentation into a set of isolated video object segmentations of
single video objects. We identify and quantify our sources of segmenta-
tion information and formulate them in terms of surface optimization.

SURFACE VS. VOLUME REPRESENTATION
As shown in Figure 4.3, we represent the video object as a 2-D surface
in the video space. Like the image segmentation problem in Chapter 2,
there are two ways to represent a video object from a graph-theoretic
perspective: either 1) volume-based as a partition of the 3-D dimensional
spaces, or 2) surface-based as a set of interlocking surfaces. Although
both exist in 3-D video space, a volume-based approach is effectively
a 3-D problem while a surface is 2-D. With optimization techniques,
the quality of the solution is often related to the number of parameters
to optimize. A surface-based description is generally simpler because
its behavior requires fewer parameters. Furthermore, by decomposing