Three dimensional interpretation of an indoor scene from a single image

One of the main goals of computer vision has been understanding the overall scene and the various components in the scene from an image. In order to guarantee that the resulting output is valid in three dimensions, three dimensional reasoning must be done. My previous work on reconstruction of indoor building structures and detecting cuboid and rectangular objects have made use of strict three dimensional reasoning, inspired by line drawing labeling work. It was able to successfully detect these components of an indoor scene and the resulting estimate was guaranteed to be physically valid due to the geometrical reasoning. I propose to combine the outputs of my existing system to build an integrated framework for understanding various components of an indoor environment with joint reasoning about the geometric relationships among the components. The benefit of joint estimate will be twofold: knowledge of one object will aid in finding other objects, and geometric reasoning will rule out incompatible objects so the resulting interpretation of a scene is geometrically coherent.