Grouping uncertain oriented projective geometric entities with application to automatic building reconstruction

Grouping Uncertain Oriented Projective Geometric Entities with Application to Automatic Building Reconstruction The fully automatic reconstruction of 3d scenes from a set of 2d images has always been a key issue in photogrammetry and computer vision and has not been solved satisfactory so far. Most of the current approaches match features between the images based on radiometric cues followed by a reconstruction using the image geometry. The motivation for this work is the conjecture that in the presence of highly redundant data it should be possible to recover the scene structure by grouping together geometric primitives in a bottom-up manner. Oriented projective geometry will be used throughout this work, which allows to represent geometric primitives, such as points, lines and planes in 2d and 3d space as well as projective cameras, together with their uncertainty. The first major contribution of the work is the use of uncertain oriented projective geometry, rather than uncertain projective geometry, that enables the representation of more complex compound entities, such as line segments and polygons in 2d and 3d space as well as 2d edgels and 3d facets. Within the uncertain oriented projective framework a procedure is developed, which allows to test pairwise relations between the various uncertain oriented projective entities. Again, the novelty lies in the possibility to check relations between the novel compound entities. The second major contribution of the work is the development of a data structure, specifically designed to enable performing the tests between large numbers of entities in an efficient manner. Being able to efficiently test relations between the geometric entities, a framework for grouping those entities together is developed. Various different grouping methods are discussed. The third major contribution of this work is the development of a novel grouping method that by analyzing the entropy change incurred by incrementally adding observations into an estimation is able to balance efficiency against robustness in order to achieve better grouping results. Finally the applicability of the proposed representations, tests and grouping methods for the task of purely geometry based building reconstruction from oriented aerial images is demonstrated. It will be shown that in the presence of highly redundant datasets it is possible to achieve reasonable reconstruction results by grouping together geometric primitives.