Multi-view depth estimation

Abstract

It is often desired to have a digital representation of a real world scene. One approach to retrieve digital information about a scene is to capture images from it by means of digital cameras. The problem with this method is that only a two dimensional representation of the three dimensional scene is available and limits us to completely reconstruct it. The problem that is being solved by multi-view depth estimation algorithms is the problem of retrieving information about the third dimension, depth. When the depth estimation for each pixel is correct we can digitally reconstruct the scene in three dimensions. Depth estimation algorithms can work on different input parameters. Some algorithms use only images from one point-of-view while other algorithms use captured data from different viewpoints. In this thesis three different types of algorithms will be discussed. First a depth from defocus algorithm is presented that uses images from only one pointof view. This is followed by explaining the general design of stereo view depth estimation algorithms and their input and output data. Later in the thesis an implementation of both types is described. The last type of algorithm that is discussed uses images from more than two different viewpoints to derive depth for pixels in the reference images.