Generating 3D Models from 2D Images

Abstract

In this Final Year Project we attempt to regain depth information from two images of a scene. Recognising depth from images is a typical problem in computer vision and a very active research area. A main reason towards the popularity of this problem is the large number of applications which could be enhanced by extracting depth information. These include automobile driver assistance, traffic management, uses in architecture, uses for entertainment purposes, image analysis and uses in robotics. To retrieve depth from the two input images, they will be formulated as a Markov Random Field using Bayesian theory. The Belief Propagation algorithm will be used to minimize the energy in the formulated Markov Random Field. This technique has shown to be effective in modem approaches to stereo matching. An important aim in this Final Year Project was to make no restrictions on the camera angles and motion, unlike other systems which require exact parameters of the camera. This introduces a new issue to our Final Year Project; before executing Belief Propagation, we must correctly rectify the images. Image Rectification will be studied and a planar rectification technique chosen to prepare the input for the Belief Propagation algorithm. The output of the system will take the form of a red-blue anaglyph image which can be viewed using standard red-blue glasses.