Stereo Reconstruction of Building Interiors with a Vertical Structure Prior

Abstract

Image-based computation of a 3D map for an indoor environment is a very challenging task, but also a useful step for vision-based navigation and path planning for autonomous systems, and for efficient visualization of interior spaces. Since computational stereo is a highly ill-posed problem for the typically weakly textured, specular, and even sometimes transparent indoor environments, one has to incorporate very strong prior assumptions on the observed geometry. A natural assumption for building interiors is that open space is bounded (i) by parallel ground and ceiling planes, and (ii) by vertical (not necessarily orthogonal) wall elements. We employ this assumption as a strong prior in dense depth estimation from stereo images. The additional assumption of smooth vertical elements allows our approach to fill in plausible extensions of e.g. walls in case of (non-vertical) occlusions. It is also possible to explicitly detect non-vertical regions in the images, and to revert to more general stereo methods only in those areas. We demonstrate our method on several challenging stereo images of office environments.