A Multiple Graph Cut Based Approach for Stereo Analysis Ulas Vural and Yusuf Sinan Akgul GIT Vision Lab, Department Of Computer Engineering Gebze Institute Of Technology, Cayirova, Gebze, Kocaeli 41400, Turkey {uvural, akgul}@bilmuh.gyte.edu.tr Abstract. This paper presents an optimization framework for the 3D reconstruction of the surfaces from stereo image pairs. The method is based on employing popular graph cut methods under the dual mesh op- timization technique. The constructed system produces noticeably better results by running two separate optimization processes that communi- cate with each other. The communication mechanism makes our system more robust against local minima and it produces extra side informa- tion about the scene such as the unreliable image sections. We validated our system by running experiments on real data with ground truth and we compared our results with the other optimization methods, which showed the accuracy and effectiveness of our method. 1 Introduction The classical breakdown of 3D surface recovery from stereo suggests that first the correspondences between the image pairs should be established and then the 3D surface is reconstructed using these correspondences[9]. The newer techniques take the approach of a global solution by incorporating the correspondence and the 3D reconstruction steps into the same process. This process is larger and more complex but the results are far better than the classical methods if the problem complexity is addressed properly. One common method to manage this larger problem is to pose it as an energy optimization task. An energy functional that penalizes locally unsmooth and discontinuous 3D structure is formulated. Optimization of this func- tional on the stereo image pairs would produce the desired 3D surface. Despite the elegance and unified nature of such systems, optimizing these functionals are not trivial. The problem is fundamentally NP-Hard and the approximation methods are sensitive to initializations, local minima, and image noise. Recently, graph cut methods gained popularity in optimizing energy function- als of Computer Vision problems. Graph cuts can guarantee optimal functional values for some restricted cases[10][8]. For the other cases, they guarantee an up- per bound in error from the optimal result[6]. Furthermore, since they are based on the deep theory of graph and flow algorithms, there are numerically stable and efficient algorithms for performing cuts[4]. Although the types of energy func- tionals that can be optimized by graph cuts are limited[13][7], the limitations are K. Franke et al. (Eds.): DAGM 2006, LNCS 4174, pp. 677–687, 2006. c  Springer-Verlag Berlin Heidelberg 2006