[Colloquium] REMINDER: 1/29 Research at TTIC: Ayan Chakrabarti, TTIC

[Mary Marre]

When:     Friday, January 29th at noon

Where:    TTIC, 6045 S Kenwood Avenue, 5th Floor, Room 526

Who:      Ayan Chakrabarti, TTIC


Title: Architectures for Learning in Low-level Vision Applications

Abstract:
Low-level vision is the processing of images and visual data to recover
physical scene attributes---shape, color, material, motion, etc.---that are
useful for a broad set of tasks. However, the mapping from scene properties
to images is many to one, and recovering the former essentially requires
solving an ill-posed inverse problem. We seek to develop data-driven
approaches to learn estimators that can sift through these many possible
scene explanations for an image and arrive at the right one, relying on the
statistics of natural scenes to eliminate answers that are unlikely. To do
so successfully, we must choose the right training strategy and parametric
form for these estimators, in-part motivated by physical models of image
formation.

In this talk, I will discuss recent work on developing learning-based
methods for two low-level vision applications. The first is computational
color constancy---recovering the true color of objects from observed color,
by estimating and correcting for the effect of unknown scene lighting. I
will present a new method that only uses the pixel-wise statistics of color
images without any spatial or semantic context, but is able to outperform
state-of-the-art color constancy methods based on complex global features.
Our approach is based on encoding these pixel-wise statistics in the form
of conditional luminance-chromaticity histograms. I will describe how these
histograms are used to extract distributions over illuminant color from
each pixel---which are then simply averaged across pixels to yield a global
illuminant estimate---as well as how we learn these histogram end-to-end
using stochastic gradient descent.

In the second part of the talk, I will present a new method for blind
motion deblurring. At the core of the method is a deep neural network that,
given an observed image patch blurred by an unknown motion kernel, computes
an estimate of its latent sharp version. The network's architecture is
motivated by a signal-processing view of image statistics and blur
analysis. It takes as input a parsimonious encoding of the observed patch
as coefficients of a multi-resolution decomposition, and has initial layers
whose connectivity is limited by locality in frequency. Moreover, instead
of directly predicting intensities, the network outputs the complex Fourier
coefficients of a restoration filter. During inference, the network is
independently applied to every overlapping patch in an observed image, and
the phase values of filter Fourier coefficients are harmonized for local
coherence. Then, overlapping filtered patches are averaged to form an
estimate of the latent sharp image, from which we estimate a global motion
kernel estimate. This method matches the performance of state-of-the-art
iterative algorithms for motion deblurring---at a fraction of their
computational cost.


********************************************************************

*Research at TTIC Seminar Series*

TTIC is hosting a weekly seminar series presenting the research currently
underway at the Institute. Every week a different TTIC faculty member will
present their research.  The lectures are intended both for students
seeking research topics and adviser, and for the general TTIC and
University of Chicago communities interested in hearing what their
colleagues are up to.

To receive announcements about the seminar series, please subscribe to the
mailing list: https://groups.google.com/a/ttic.edu/group/talks/subscribe

Speaker details can be found at: http://www.ttic.edu/tticseminar.php.

For additional questions, please contact David McAllester at
mcallester at ttic.edu



Mary C. Marre
Administrative Assistant
*Toyota Technological Institute*
*6045 S. Kenwood Avenue*
*Room 504*
*Chicago, IL  60637*
*p:(773) 834-1757*
*f: (773) 357-6970*
*mmarre at ttic.edu <mmarre at ttic.edu>*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20160128/73a490d3/attachment-0001.htm