Segmenting multiple range images with primitive shapes

We introduce a novel method for automatically segmenting multiple registered range images by detecting and optimizing geometric primitives. The resulting shapes provide high level information about scanned objects and are a valuable input for surface reconstruction, hole filling, or shape analysis. We begin by generating a global graph of sample points covering all input frames. The graph structure allows to compute a globally consistent segmentation with a memory and time-efficient solution, even for large sets of input images. We iteratively detect shapes with a Ransac-approach, optimize the assignments of graph nodes to shapes, and optimize the shape parameters. Finally, pixel-accurate segmentations can be extracted for each source image individually. By using range images instead of unstructured point clouds as input, we can exploit additional information such as connectivity or varying precision of depth measurements.