COLLOQUIUM
             Computer Science Department, Boston University

Speaker: Stella X. Yu
         Computer Science Division,
         UC Berkeley

Date:  Thursday, March 24
Time:  12:30
Place: Room MCS 135, 111 Cummington Street 
       (for directions, see www.cs.bu.edu/colloquium)

Title: Towards Understanding Visual World from Real Images

Abstract:

To understand the visual world depicted in a real image involves both
recognizing what objects are in the image and how they are segmented from
the surrounding background.  My research explores the computational
underpinnings that underlie such a seemingly effortless perceptual ability.

In this talk, I will focus on two recent projects:
1. object segmentation as a concurrent grouping problem, and
2. general-purpose image segmentation by scale invariance.

Object segmentation is a typical chicken-and-egg problem.  My work treats
it as an optimal grouping problem at both pixel level and object-part
level, so that segmentation and recognition can be realized efficiently in
a concurrent fashion.

The main challenge in segmentation is to be able to discover region
boundaries regardless of contour gaps and complex textural appearance.  By
optimizing a consistent grouping over scales, I will demonstrate that
segmenting images of general complexity is possible based on entirely
multiscale edges without resorting to any explicit modelling of texture or
curvilinearity.

My work in computer vision has contributed to solving semi-supervised data
clustering problems, and continues to provide new ideas to large-scale data
mining problems.  It has also produced new computational tools that have
already found their use in computer graphics, computational geometry, and
protein identification.

Biography:

Stella X. Yu obtained a B.S. in Information Science and Technology from
Xi'an Jiaotong University (China), an M.S. in Pattern Recognition and
Intelligent Control from Tsinghua University (China), and a Ph.D. in
Robotics from School of Computer Science at Carnegie Mellon University,
with completion of graduate training program in the Center for the Neural
Basis of Cognition at Pittsburgh.  Since May 2003, she has been a postdoc
in the Computer Science Division at UC Berkeley.  Her main research
interests are computer vision, computational modelling of human vision, and
machine learning.

Host: Margrit Betke