Maximum Likelihood Features for Generative Image Models

• Activity Number: 514 - Recent Advances in Imaging Statistics: Bayesian Methods and Beyond
• Type: Contributed
• Date/Time: Wednesday, August 2, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Lo-Bin Chang* and Eran Borenstein and Wei Zhang and Stuart Geman
• Companies: The Ohio State University and Microsoft Research and Smartleaf, Cambridge and Brown University
• Keywords: generative models ; conditional modeling ; appearance models ; image features ; image models ; computer vision

Abstract

A forward or generative model in Bayesian image analysis consists of a prior probability distribution and a conditional data model. The prior distribution expresses common or learned knowledge about the odds of possible interpretations; the conditional data model places a probability on image-based observations given any particular interpretation. To the extent that the generative model generates features, as opposed to pixel intensities, the posterior distribution (i.e. the conditional distribution on part and object labels given the image) is based on incomplete information; feature vectors are generally insufficient to recover the original pixel intensities. Therefore, it is of interest to develop a class of data models that generate pixel intensities rather than image features. In this talk, I will propose a new and general approach to this challenge, based on which the pixel-level models for the appearances of parts and objects are built. A series of experiments in image sampling and image classification will be discussed to demonstrate the utility of this approach.