Recovering the shape of objects in 3D point clouds with partial occlusions

In this paper we present an approach to label data points in 3d range scans and to use these labels to learn prototypical representations of objects. Our approach uses associative Markov networks (AMNs) to calculate the labels and a clustering operation to determine the prototypes of homogeneously labeled regions. These prototypes are then used to replace the original regions. In this way, we obtain more accurate models and additionally are able to recover the structure of partially occluded objects. Our approach has been implemented and evaluated on 3d data of a building acquired with a mobile robot. The experimental results demonstrate that our algorithm can robustly identify objects with the same shape and can use the prototypes of these objects for highly accurate mesh completion in case of occlusions.