Monday, May 10, 2010
10:00 AM
CSB 703
Ph.D. Thesis Proposal
Chetan Bhole
University of Rochester
Conditional Random Fields for segmentation in Image Sequences
Graphical models have emerged as powerful tools for representing complex probability distributions over large groups of random variables. The discriminative type of random fields called Conditional Random Fields (CRF) model a joint conditional probability distibution on structured data. Image
segmentation is a fundamental problem in computer vision. Image sequences like volumetric medical images and spatio-temporal video have become extremely commonly available data. We have used CRFs to model three dimensional medical images and propose to extend the work to video data.
Segmenting videos can be challenging because we need to deal with motion-blur, occlusion, light and appearance changes of objects and compression artifacts not to mention the possibility of camera, foreground and background motions. More over, obtaining segmented labeled data for videos is tedious and time-consuming. We propose to handle these issues in a CRF framework.