INTRODUCTION TO CONTENT-BASED VISUAL PROCESSING - Video Object Extraction and Representation: Theory and Applications - page 21

contour, we can order the space w.r.t. an initial contour estimate,
enabling our surface optimization algorithm and integrating a
concept of shape similarity into our surface optimization.
2. Formulation of Video Object Segmentation as a Surface Opti-
mization Problem
To aid in the analysis of the video object segmentation problem,
this topic formalizes the video object segmentation problem as a
surface optimization problem. This formalization allows for the
integration of many different sources (visual, motion, knowledge-
based) of video information into a single computational frame-
work.
3. Iterative Viterbi Algorithm
Using the Voronoi Ordered Space of Chapter 3 to describe a priori
shape constraints, our novel Iterative Viterbi algorithm optimizes
a surface by using a dynamic programming algorithm within an
iterative framework. For 2-D surfaces that are highly asymmetric
in one dimension, our novel Iterative Viterbi algorithm decom-
poses the surface along the asymmetric dimension (for instance,
time in video sequences) into a series of contour optimizations
that can be optimally solved, then reassembles the contour es-
timates into a surface estimate, and recalculates each contour
optimization to reflect the continuity of surface.
4. A Novel Video Object Segmentation System
Putting together the formulation of video object segmentation
with the Iterative Viterbi algorithm, we produce a system to ex-
tract a video object segmentation.
Video Object Representation Techniques
1. Voronoi Ordered Space (Second Application)
By operating upon the Voronoi Order Space, we extract a skeleton-
like representation of a shape based upon assumptions of object
structure.
2. DAG-Coding and DAG-Compare
Our work in DAG S gives a general methodology for representing
the partitioned data [Lin and Kung, 1997a] [Lin and Kung, 1997b]
[Lin and Kung, 1998b]. We provide an algorithm for efficient
comparisons of these DAG structures. This data structure and
its comparison algorithm allow for a robust divide and conquer
method and form the basis for our shape query system.
3. DAG-Ordered Trees
From this work in Chapter 5, we also propose a new data structure