Efficient Reconstruction of Complex 3-D Scenes from Incomplete RGB-D Data

In this paper we develop a new approach for reconstructing 3-D scenes from RGB-D data. We use a Markov random field to model appearance relations and geometric cues between different regions of a scene, as a means to provide robustness to noisy and incomplete data often generated by RGB-D devices. A parametric reconstruction of 3-D scenes that enable coherent physical interaction are computed, in near real time, with a standard computer that does not use specialized hardware.