Estimating accurate depth maps by solving the stereo problem is an important step in reconstructing real world surfaces. Variational methods that minimize a global energy functional are considered as especially precise throughout the literature. By using the depth as parameterization directly, the approach is easily extended to multiple views, allowing to considerably enhance the quality of the resulting depth maps. Further improvement can be achieved by adapting the depth parameterization to the regularizer used. In this thesis secondorder regularization is used with an inverse depth parameterization and compared against the direct depth parameterization to examine its benefits. Several extensions that improve upon the naïve multiview approach are suggested. The proposed method and its extensions are evaluated by experiments, using artificial as well as real world test scenes.