VASC Seminar Announcement
       =========================

 Date: Monday April 29th
 Time: 3:30 - 4:40
 Location: NSH 1507

 Speaker:  Stella X. Yu
 Affiliation: Carnegie Mellon University, Robotics Institute and Center for the Neural Basis of Cognition
 
 Title:  Attentive Image Segmentation

Abstract:

Image Segmentation is often considered a bottom-up process.  However,
without prior knowledge about the goal of image segmentation, it is
generally an ill-defined problem.  Attention is one way to constrain the
segmentation task.  I'd like to show two points in this talk.  First, it is
possible to integrate both bottom-up and top-down information in a
single grouping process.  This process needs not to be done sequentially as
most previous approaches assume.  Second, figure-ground segregation can
sometimes be obtained without specific object knowledge.

We formulate the attentive image segmentation problem in a graph
partitioning framework.  A weighted graph is built on low level cues from
the image.  The knowledge of attention is encoded in a saliency map which
defines a partial grouping solution.  Combining these two sources of
information, our model leads to a constrained eigenvalue problem, whose
global-optimal solution can be obtained using eigendecomposition.  We
demonstrate on real images and video sequences that simple priors can
greatly improve image segmentation and approach figure-ground segregation.

Joint work with Dr. Jianbo Shi.

www.cs.cmu.edu/~xingyu/research/