INTRODUCTION TO CONTENT-BASED VISUAL PROCESSING - Video Object Extraction and Representation: Theory and Applications - page 20

a video object segmentation system partitions the space of the video
sequence into video objects. In this topic, we simplify the general video
object extraction to extraction the video objects of different movement.
Our solution involves a formalization of the Video Object Segmen-
tation problem as mathematical surface optimization where the energy
function depends on visual information, motion information and object
knowledge. Within this system, Voronoi Ordered Spaces are used as
means of integrating shape information. A novel variation upon HMM
training and the Expectation-Maximization framework called Iterative
Viterbi is presented for fitting the surface to the visual artifacts within
a video sequence.
QUERY BY SHAPE FOR MPEG-7
In Chapter 6, given a MPEG-7 video object database, we would like
to query it by shape, i.e., to ask a user for an outline and search for a
VOP that has shape with a similar outline. Shape itself is special case
of non-verbal query problem that MPEG-7 framework supports. Our
own shape descriptor is one of many descriptors that would likely show
up in a interface to MPEG-7 database. The shape query problem has
two subproblems that are innately tied together: shape extraction and
robust representation/comparison.
The shape extraction is a result from the concept of Voronoi Ordered
Space; the shape representation and comparison algorithm were drawn
from general methodology inspired by our work in cursive handwriting
recognition. Our solution expands the ordered tree paradigm with the
introduction of the DAG-ordered tree (or DOT, for short) to construct
high-level representations that avoid quantization.
7. RESEARCH CONTRIBUTIONS
In this topic, we have made both theoretical and pragmatic contribu-
tions to the field of content-based analysis. On the theoretical side, we
introduced new concepts in representation, data structure, and shape
description. In addition, a number of software systems have been devel-
oped to verify the theoretical concepts.
Video Object Extraction Techniques
1. Voronoi Ordered Space (First Application)
Based upon the Voronoi cells, the Voronoi Ordered Space is a
concept that can specify a descriptive subset of curves that are
related to an initial contour. By describing a planar surface in
terms of a projection onto a contour and its distance from the