Efficient Generation of 3-D Models Out of Depth Maps

Depth reconstruction or acquisition with a 3-D camera results in a video sequence where each pixel of a frame is annotated with a depth value. We propose an approach to combine this depth information into a 3-D model of the assumedly static scene. For processing of video, we have to satisfy strong requirements on computational and storage efficiency, incremental processing and handling of uncertainty in the depth maps. Our approach is based on an octree structure. Cells are split until the topology of the scene can be recovered uniquely. A signeddistance function to the nearest surface is computed in the octree vertices. The scene can then be reconstructed by extraction and triangulation of an implicit surface. The depth uncertainty results in boundary conditions for the implicit surface.