Stereo Matching and Graph Cuts
Ayman Zureiki, Michel Devy, Raja Chatil · InTech eBooks · 2008
We described in this chapter, our evaluation of global stereo correspondence methods based on graph cuts. The combination of a local method, able to select a reduced set of possible matches for each pixel, and a global method, based on the graph cuts algorithm, let us to achieve two goals ? Sensibly ameliorate the quality of disparity images obtained only by a local method. ? Avoid the combinatorial explosion of the Graph Cuts method executed without preliminary reduction of the graph. Some optimizations of our algorithm are currently studied. Note that we work always on pre-rectified images, and we produce an integer disparity image: we will study how to adapt this algorithm in order to find stereo matching on non-rectified images. The reduced graph approach allows us to limit the execution time and to use larger disparity ranges. We write the code with the aim to obtain results and not to optimize the performance. In addition, we did not use the best maximum flow algorithm. In the next period, we could very easily improve performances, in order to satisfy as much as possible, real time constraints. Section 4.4 has shown that the reduced graph is able to deal with sub-pixel disparities (real values). Testing this property and obtaining sub-pixel disparity images will be our main objective in our near future works.