Digital three-dimensional (3D) models are of significant interest to many application fields, such as medicine, engineering, simulation, and entertainment. Manual creation of 3D models is extremely time-consuming and data acquisition, e.g., through laser sensors, is expensive. In contrast, images captured by cameras mean cheap acquisition and high availability. Significant progress in the field of computer vision already allows for automatic 3D reconstruction using images. Nevertheless, many problems still exist, particularly for big sets of large images. In addition to the complex formulation necessary to solve an ill-posed problem, one has to manage extremely large amounts of data. This thesis targets 3D surface reconstruction using image sets, especially for large-scale, but also for high-accuracy applications. To this end, a processing chain for dense scalable 3D surface reconstruction using large image sets is defined consisting of image registration, disparity estimation, disparity map fusion, and triangulation of point clouds. The main focus of this thesis lies on the fusion and filtering of disparity maps, obtained by Semi-Global Matching, to create accurate 3D point clouds. For unlimited scalability, a Divide and Conquer method is presented that allows for parallel processing of subspaces of the 3D reconstruction space. The method for fusing disparity maps employs local optimization of spatial data. By this means, it avoids complex fusion strategies when merging subspaces. Although the focus is on scalable reconstruction, a high surface quality is obtained by several extensions to state-of-the-art local optimization methods. To this end, the seminal local volumetric optimization method by Curless and Levoy (1996) is interpreted from a probabilistic perspective. From this perspective, the method is extended through Bayesian fusion of spatial measurements with Gaussian uncertainty. Additionally to the generation of an optimal surface, this probabilistic perspective allows for the estimation of surface probabilities. They are used for filtering outliers in 3D space by means of geometric consistency checks. A further improvement of the quality is obtained based on the analysis of the disparity uncertainty. To this end, Total Variation (TV)-based feature classes are defined that are highly correlated with the disparity uncertainty. The correlation function is learned from ground-truth data by means of an Expectation Maximization (EM) approach. Because of the consideration of a statistically estimated disparity error in a probabilistic framework for fusion of spatial data, this can be regarded as a stochastic fusion of disparity maps. In addition, the influence of image registration and polygonization for volumetric fusion is analyzed and used to extend the method.
Finally, a multi-resolution strategy is presented that allows for the generation of surfaces from spatial data with a largely varying quality. This method extends state-of-the-art methods by considering the spatial uncertainty of 3D points from stereo data. The evaluation of several well-known and novel datasets demonstrates the potential of the scalable stochastic fusion method. The strength and the weakness of the method are discussed and direction for future research is given.
«Digital three-dimensional (3D) models are of significant interest to many application fields, such as medicine, engineering, simulation, and entertainment. Manual creation of 3D models is extremely time-consuming and data acquisition, e.g., through laser sensors, is expensive. In contrast, images captured by cameras mean cheap acquisition and high availability. Significant progress in the field of computer vision already allows for automatic 3D reconstruction using images. Nevertheless, many pro...
»