INC Seminar Series
(Beverages & Desert Provided)

Tuesday,  April 9, 2002   NOON
Cognitive Science Building
Room 003

Joachim Denzler
Chair for Pattern
Recognition 
University of Erlangen, Germany
Optimal Active Vision for Object Recognition
--------------------------------------------

The term active vision has been used in the past
to emphasize the differences to the passive
or Marr appoach of vision. The term "active" does not just mean to
see but to look. From the computer vision point of view
this has been understood mostly as actively acquiring sensor
data and selectively processing it.

In the talk an important problem in computer vision is tackled
from the active vision point of view: object recognition.
Recognition can be improved by providing the classifier
with the right sensor data, i.e. avoiding ambiguities in the taken
sensor information. Avoiding ambiguities is approached not
in the learning phase of a classifier but during recognition itself.

Two methods are presented for selecting the
optimal sensor data: one based on reinforcement learning, the
other using information theoretic concepts. Experiments on optimal
viewpoint selection as well as on optimal gaze control show
the benefits of an active strategy compared to a passive one.