CS Seminar: Jianbo Shi, Wed. 6 December

[Margery Ishmael]

CS SEMINAR ANNOUNCEMENT:

Wednesday, 6 December at 2:30 pm in Ryerson (Annex) 276
(To be followed by refreshments in Ryerson 255)

Jianbo Shi
Robotics Institute
Carnegie Mellon University

Title: Image Segmentation: finding global pattern through learning and 
integration of pairwise proximity and repulsion information

Abstract:
In this talk I will describe our recent work on grouping and image 
segmentation in two directions:
1) learning segmentation from examples, and
2) segment data with proximity as well as repulsion information.
This work builds upon our initial Normalized Cuts segmentation
algorithm, which formulates segmentation as a hierarchical graph
partitioning problem. The normalized cut criteria is a global balanced
measure which minimizes similarity between the segments, and
maximizes similarity within the segments simultaneously. An
efficient computation algorithm is developed by solving a sparse
generalized eigen-equation.

We have demonstrated our segmentation algorithm on 1000 natural images
using simple lower level visual cues such as texture and contour.
However, in human perception higher level object knowledge, such as
shape, pays an important role in segmentation. We propose a solution
for learning segmentation of a specific object through training
examples. The solution is based on exploring a tight connection
between Markov random walk on a graph and the normalized cut
algorithm. By "tuning'' the local random walk transition probability
distribution, the desired global segmentation result can be achieved.
In the second direction, we have generalized the Normalized Cuts
algorithm for integrating pairwise proximity as well as repulsion
information. This algorithm provides us a greater control the
segmentation process, while reducing the number of necessary pairwise
connections. We have tested this algorithm on Figure-Ground image
segmentations by encoding figure-ground separation cue using repulsion
connections, and encoding grouping cues using proximity connections.
This is a joint work with Maria Maila and Stella Yu.

Host: Partha Niyogi
====================================================
Margery Ishmael
Department of Computer Science
1100 E. 58th St.
Chicago, IL 60637

tel: 773 834-8977  fax: 773 702-8487

marge at cs.uchicago.edu 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20001129/b30d197d/attachment.htm