[Colloquium] Reminder: Larsson/MS Presentation/Apr 8, 2014

[Margaret Jaffey]

This is a reminder about Gustav's MS Presentation tomorrow.

------------------------------------------------------------------------------
Date:  Tuesday, April 8, 2014

Time:  3:30 PM

Place:  Ryerson 277

M.S. Candidate:  Gustav Larsson

M.S. Paper Title: Object detection with CAD-trained statistical models

Abstract:
Using 3D CAD models to train computer vision systems is an attractive
premise that reconnects us with earlier work within the field. The
prospects of an ample source of noiseless and automatically
pose-annotated training data sounds too good to be true. Perhaps it
is, since so far attempts made by the community have not been close to
matching those of natural data, making the supposed benefits moot.

In this work, we argue that the performance discrepancy is not
necessarily due to the difference in image statistics. Instead, one of
the main problems is that objects in natural images follow a
heterogeneous camera pose distribution, favoring certain angles.
Training from data where this distribution naturally occurs, infuses
the model with a strong prior that will be beneficial at testing time.
However, current attempts of using CAD models render the objects from
a uniform camera pose distribution, within a certain set of
constraints. By selecting a camera pose distribution with only a few
hand-picked nonzero points, we show that detection rates are greatly
improved over the uniform alternative, even with much less training
data.

The paper also explores using simple generative models for building
the detector. Algorithmically, our models are trivial compared to our
baseline of choice, the linear SVM. However, that does not mean the
problems we face are trivial, such as how to deal with correlated
features when the model assumes independence. In a sense, these
problems are solved automatically by the SVM. We believe that exposing
them and trying to solve them explicitly can be a great source of
knowledge and eventually lead to better models. By carefully
addressing some of these problems, we show that we can achieve results
not far behind the baseline.

In the same vein, we do not use any natural images as part of the
training process, even though methods that mix synthetic and natural
training data have been fairly successful. Training solely from CAD
data forces us to face the problems associated with this approach and
exposes the real reasons behind the discrepancy in error rates.

Gustav's advisor is Prof. Yali Amit

Login to the Computer Science Department website for details:
 https://www.cs.uchicago.edu/phd/ms_announcements#larsson

=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=
Margaret P. Jaffey            margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 156)               (773) 702-6011
The University of Chicago      http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=