Jobs & Karriere

In classical dense stereo matching, the result of two or more 2D input images is a dense point cloud in 3D space. For practical reasons, one often needs to mesh the point cloud to a triangulated isosurface, which then can be textured to produce visually appealing results. The meshing of point clouds can be described as fitting a closed surface, parametrized by e.g. triangles, through the input points. For dense and uniformly sampled point clouds without outliers, this task can be done quite well using established algorithms like for example “Marching Cubes” or “Poisson Surface Reconstruction”.

But when obtaining a 3D point cloud from remote sensing images via dense stereo matching, we face three major problems:

1) Especially in urban areas, normally one or two sides of a building are missing due to occlusion.

2) When merging different point clouds (of the same scene) obtained via multi-view matching, the point cloud normally is quite dense, but at the same time not uniformly sampled and (due to errors in the image registration) prone to some noise. Therefore a 3D point in the real scene often is represented by many representatives slightly different in position.

3) The initial point clouds are far to dense for large-scale applications. A standard aerial stereo image-pair for example results in roughly 30 million points or 60 million triangles for a square kilometer. Fitting a triangulated isosurface to these point clouds is challenging for both of the two first cases. For the third case, the goal is to exploit locally planar structures in the (noisy) point clouds to reduce the input data amount by >95%, while maintaining the same visual quality (detect locally planar surfaces and reduce/re-arrange the mesh).

Tasks:

• Getting an overview of the State-of-the-Art algorithms
• Choosing the best fitting algorithms and implementing them into the existing framework
• Evaluating the accuracy of the algorithms on ground truth data