Abstract: Techniques are provided for modeling indoor scenes including receiving a request for a 3D model of an indoor scene based on multiple flat images of the indoor scene, where the images obey no more than a limited overlap requirement, are absent of depth information, and are taken from one or more viewpoints. The techniques proceed by determining vanishing points in the images, receiving floor contour information that was determined based on the vanishing points; reconstructing the 3D vertex positions of two or more floor plan parts using a geometric constraint matrix that encodes coordinate equalities among said vertices, based on the floor contour information; and assembling a combined floor plan based at least in part on the floor plan parts. The techniques proceed by receiving a floor plan outline indicating walls and generating the 3D model of the indoor scene based on the floor plan outline.