2 1/2D scene reconstruction of indoor scenes from single RGB-D images

Abstract

Using the Manhattan world assumption we propose a new method for global 2 1/2D geometry estimation of indoor environments from single low quality RGB-D images. This method exploits both color and depth information at the same time and allows to obtain a full representation of an indoor scene from only a single shot of the Kinect sensor. The main novelty of our proposal is that it allows estimating geometry of a whole environment from a single Kinect RGB-D image and does not rely on complex optimization methods. This method performs robustly even in the conditions of low resolution, significant depth distortion, nonlinearity of depth accuracy and presence of noise. © 2013 Springer-Verlag.