Abstract
The reconstruction of a scene is a fundamental task in many AR applications. Only a dense and complete surface representation enables realistic AR interactions, like occlusions and collisions. The reconstruction of arbitrary scenes is an important issue and many approaches have already been proposed. The results of these methods are mostly dense and accurate, but the reconstructions are still incomplete. This is due to the fact that parts of the scene are invisible during the capturing process (e.g. occluded by another object). We address the problem of filling unknown parts of a surface reconstruction during the capturing process by expanding the well-known KinectFusion pipeline. Our approach diffuses the observed parts to extend the original reconstruction.By showing the results of several scenes, we demonstrate that our algorithm produces visually plausible surfaces. Moreover, we present a method for handling of occlusions and collisions directly on the extended surface.