Abstract:

Correspondence is the task of matching elements from one set to those
of another, and arises in varied scenarios such as language
understanding, protein structure prediction, and stereo image
alignment. Correspondence of data to a model, while often
computationally challenging, offers the benefit of a consistent
representational space. Such a semantic representation can allow us to
answer some questions more naturally and easily, such as whether a
cheetah is running, or whether the vibration of a vocal chord is
pathological.

I will first formulate the correspondence problem as constrained
matching that can be solved via probabilistic inference in a Markov
random field, and present two methods for coping with the
computational difficulty. The first method combines combinatorial
optimization within a meta-belief-propagation algorithm, and takes
advantage of situations where mutual exclusion constraints exist. The
second approach uses current beliefs in order to guide the propagation
process, and leads to significant improvements in speed and
convergence in the more general case.

I will then present three image based applications where probabilistic
model correspondence is particularly useful. First, I will show how it
allows us to accurately align microscopy images with low SNR,
facilitating 3D cell reconstruction. This allowed us to take the human
expert out of the loop for a wide range of challenging datasets,
without compromising the quality of the reconstruction. Second, I will
demonstrate how corresponding an object class model to an image allows
us to naturally answer semantic queries such as whether a cheetah is
running or whether an airplane is taking off. Importantly, I will show
that we are able to answer these questions without supervision at
training time, while still being competitive with a fully supervised
discriminative approach. Finally, I will describe an ongoing work for
the analysis of vocal chords pathologies from videos. By making use of
active contours that are designed to promote correspondence, we are
able to track vocal chords even where standard segmentation cannot
succeed (\eg when the chords are closed). This allows us to analyze
the full motion of the chords and, for the first time, automatically
localize pathologies.