|
Schemes for Engineers in Research and Development
Global Research Awards: Profiles
Dr Peter Hall – University of Bath
Understanding Dynamic Textures
Dynamic textures are complicated class visual
phenomena that are little understood by the computer
vision community. The class is defined by a
more-or-less arbitrary list that is not limited to
water, fire, trees, clouds, crowds, and flags. In
fact, the only property that seems to bind these
objects is that they are not rigid or articulated
bodies and their motion is highly complex.
The purpose of the sabbatical period was to find a
way to describe dynamic textures so that users can
interact with them. For example, a Hollywood film
producer may want to control an avalanche, or a
computer games company may like to include fire, or
a teenager may want to create animations on their
desktop. Each of these requires different user
tools, they have vastly differing budgets, and
diverse objectives. But the common link is that each
requires a description of dynamic texture that must
separate appearance from motion.
Our hosts, the Intelligent Systems Laboratory at the
University of Amsterdam (UvA) are acknowledges
experts in the field of image statistics, and in
addition have an interest in perception. Their
experience complemented our own, in computer vision
applied to computer graphics, in a neat and useful
way.
During the sabbatical period we investigated several
alternatives to describe shape; we imagined a video
as a collection of changing shapes. We found
investigated several routes; histograms of distance
transforms and Radon transforms showed promise. We
also looked at using the “principle of stability” ,
in which a description should be change little, if
the video content is more-or-less the same. We found
the principle can act to unify spatial segmentations
as well as underpin temporal coherence.
At the same time we worked on a special case example
of dynamic textures: trees. It was shown that a
hierarchical tracker produces the most stable
description of trees in video, separates appearance
from motion and can be used to construct a
probabilistic generative model.
In sum, we completed a special case example that
provides a sound direction for future work – and
have a much more informed view of how to generalize
away from trees to the wider class of dynamic
textures. The sabbatical period has resulted in a
grant application being written, and several papers
are planned. In addition, we found scope for our
students to work together – even some undergraduate
work at Bath found its way into the labs at UvA.
Both UvA and ourselves feel the collaboration was
worthwhile, and we wish to continue in the future.
We both wish to thank to Royal Academy of
Engineering for providing us with the opportunity.
Dr Peter Hall
Media Technology Research Centre
Department of Computer Science
University of Bath
|
