Single view object completion

Abstract

The goal of this paper is to achieve complete reconstruction of a 3d object from a single depth image observation. Much effort has been put on multi-view reconstruction of objects with substantial success, but single view recon- struction is still very limited. Initial methods produce only partial reconstructions by projecting a depth map. State of the art approaches achieve complete reconstruction but either require user interaction or perform successfully on only a handful of simple categories. The method described in this paper is an exemplar based approach to fully au- tomated reconstruction of a large variety of object classes using only a single depth image. The approach has three major components: retrieving a similar object, fitting the matched object to the query point cloud using alignment and symmetries, and reconstructing the mesh using the exemplar as a template. This method is evaluated in three distinct experiments: novel category (query of untrained class), novel model (query of trained class, untrained model), and novel view (query of trained model from a new viewpoint).