Abstract

Building a complete 3D model of a scene, given only asingle depth image, is underconstrained. To gain a full vol-umetric model, one needs either multiple views, or a singleview together with a library of unambiguous 3D models thatwill fit the shape of each individual object in the scene. We hypothesize that objects of dissimilar semanticclasses often share similar 3D shape components, enablinga limited dataset to model the shape of a wide range of ob-jects, and hence estimate their hidden geometry. Exploringthis hypothesis, we propose an algorithm that can completethe unobserved geometry of tabletop-sized objects, based ona supervised model trained on already available volumetricelements. Our model maps from a local observation in asingle depth image to an estimate of the surface shape inthe surrounding neighborhood. We validate our approachboth qualitatively and quantitatively on a range of indoorobject collections and challenging real scenes.