The Use of Bayesian Networks for Structured 3-D Object Descriptions

ZuWhan Kim


Abstract

Computing 3-D object level descriptions from images is a key goal of computer vision. Hypothesize and verify paradigm is a common approach to accomplish this goal. A key problem here is evaluation of a hypothesis based on evidence that is uncertain. There have been few efforts on applying formal reasoning methods on this problem. We propose a use of Bayesian network on this problem and show an experiment with a multi-view building detection system. There have been few efforts to apply a formal reasoning method in the area. Since the standard Bayesian network does not support for evidence variables of an arbitrary number, we propose a dynamic Bayesian network (DBN) which dynamically instanciates its structure according to the input evidence. The structure of the Bayesian network is designed by performing a correlation analysis, which brings a new concept of using hidden nodes in Bayesian network. Experimental results show that the proposed method gives good results.

On-line references


Maintained by Alexandre R.J. FRANÇOIS