Stereo korespondence

Abstract

This thesis describes how to solve the problem of stereo correspondence. This can be simplified for finding matching points in the recordset 3D scene. Searched result is the relative distance of these two points is known as the disparity. If you have enough data then with this value you can easily calculate the position of the observed points towards the observer. Thesis briefly summarizes the history of the development and research in the field of 3D scene reconstruction and closely related sectors. Mainly, however, describes the basic principles and methods of separation of different types of algorithms developed for this purpose. In particular, it analyzes the theory necessary for understanding of global Belief Propagation algorithm and possible ways to optimize this algorithm. Optimalization is implemented using threads and SSE CPU instructions. But others ways to optimize this algoritm are also mentioned as CUDA or OpenCL.