Abstract: In this talk we develop a new curve evolution formulation for
estimating
the posterior distribution of objects in images. Similar to level sets,
we
describe the segmentation of images via a conventional likelihood model
combined with a curve prior on boundaries. Unlike level sets, the
curves are
encoded via the logarithm-of-odds representing the posterior
distribution on
labels in an unconstrained vector space. The posterior distributions
are sought
via the Mean Field approach. By choosing a different representation and
optimization, our framework sidesteps many of the issues traditionally
associated with level set implementations. For example, applications
with more
than two labels are easily accommodated, as the label assignment is
accomplished by the Maximum A Posteriori rule, so there are no problems
of
overlap or vacuum between level set functions. We demonstrate the
effectiveness
of this technology on a synthetic noisy image and real 3D medical scans.