Computing That Serves

For the Birds: Pose-normalized Models for Fine-grained Object Recognition


Thursday, January 10, 2013 - 11:00am


Ryan Farrell

Research Scientist, International Computer Science Institute


Bryan Morse

Colloquium in 1170 TMCB

Computer vision is quietly finding its way into our everyday lives,
playing a central role in today's cars, digital cameras, and
smartphones, among other devices.  Successful research in object
recognition over the last decade has enabled many of these new
technologies and it remains the most actively researched area within
computer vision.  Two emerging directions that are transforming the
field are large-scale and fine-grained recognition.  Real-world
recognition is inherently a big-data problem, requiring recognition of
tens of thousands of categories.  Within a given object domain (e.g.
birds, cars), the key challenge presented by fine-grained recognition is
finding the highly-localized and often subtle characteristics
(differences in shape, coloration, etc.) that allow precise
identification at the level of subordinate categories (e.g.
genus/species, make/model/year).  I will present two models that
overcome this key challenge of localizing the discriminative visual
information: the first approach uses a volumetric representation to
model appearance in a pose-normalized space; the second approach uses
semantic pooling, aggregating appearance information with respect to
predicted part locations.  I will conclude by describing a large-scale
avian dataset that my colleagues and I are creating by employing novel
human computation techniques.


Ryan Farrell is a research scientist
at the International Computer Science Institute (ICSI), a non-profit
research institute affiliated with UC Berkeley.  He joined ICSI as a
postdoctoral researcher in 2011 after completing master's and doctorate
degrees at the University of Maryland, College Park.  Ryan worked for
three years as a software engineer after completing his bachelor's
degree at UC Berkeley before returning for his graduate studies.  Ryan's
research interests include various topics in computer vision, machine
learning, and artificial intelligence.  His current work focuses on
object recognition, specifically on the problem of fine-grained visual
categorization.  He is also interested in emerging topics such as
big-data and human computation.