Abstract

Dense 3D reconstruction still remains a hard task for abroad number of object classes which are not sufficientlytextured or contain transparent and reflective parts. Shapepriors are the tool of choice when the input data itself is notdescriptive enough to get a faithful reconstruction. We pro-pose a novel shape prior formulation that splits the objectinto multiple convex parts. The reconstruction problem isposed as a volumetric multi-label segmentation. Each of thetransitions between labels is penalized with its individualanisotropic smoothness term. This powerful formulation al-lows us to represent a descriptive shape prior. For the objectclasses used in this paper the individual segments naturallycorrespond to different semantic parts of the object. Thisleads to a semantic segmentation as a side product of ourshape prior formulation. We evaluate our method on several challenging real-world datasets. Our results show that we can resolve issues such as undesired holes and discon-nected parts. Taking into account a segmentation of the freespace, we show that we are able to reconstruct concavities, such as the interior of a mug.