Seminars

Conditional Modeling for Image Analysis

66
reads

Lo-Bin Chang

2012-03-30
15:00:00 - 16:40:00

103 , Mathematics Research Center Building (ori. New Math. Bldg.)

A probabilistic framework for modeling images of objects constitutes a generative scene model of parts and a Bayesian framework for image analysis. In this talk, I will propose a conditional modeling technique to finesse the complexity of the high dimensional image data and to develop the context sensitive prior distribution on a hierarchy. I will argue that the pixel-level models for the appearances of parts can be learned using the technique, and I will show an approximate sampling formula using the projection method. In addition, a particular relationship of parts, Relational Coordinate System, for tilting the distribution will be discussed in the end of the talk.