Single View Corridor Reconstruction

We present an autonomous algorithm for 3d reconstruction from a single image of an indoor scene. Most work on 3d reconstruction from vision uses multiple images (for example, binocular vision/stereopsis), and recovers depths using triangulation. However, the range at which binocular vision is accurate is limited by the “baseline” distance between the two cameras. In contrast, humans are often able to look at a monocular (single-camera) image, and guess how far away objects were from this camera. Humans do this by bringing to bear on the problem significant prior knowledge about the world, which is used to resolve the ambiguities intrinsic to monocular vision. Focusing on the problem of 3d reconstruction in indoor scenes, in this report we present an algorithm for depth recovery based on precise segmentation of the floor in an image. Our approach encodes prior knowledge about the environment such that the floor is flat and that walls are vertical, and is based on recognizing the floor and the wall regions in the image. We also present experimental resulting showing our algorithm giving good 3d reconstruction performance using monocular images of indoor scenes.