Estimating Planar Structure in Single Images by Learning From Examples

Abstract

Outdoor urban scenes typically contain many planar surfaces, which are useful in tasks such as scene reconstruction, object recognition and navigation. Planar constraints are especially useful when only a single image is available, though the lack of 3D information makes finding them difficult; but a number of cues-such as rectangular shapes, edges, and appearance-can make this possible. We develop a method to determine if regions in an image are planar and find their orientation; motivated by how humans use their prior knowledge to help interpret new scenes, this is done by learning from a training set of examples. In contrast to previous methods which often rely on rectangular structure, this allows our method to generalise to a variety of outdoor environments, without relying on restrictive assumptions such as a Manhattan-like world or a camera aligned with the ground plane. From only one image, our method is able to reliably distinguish planes from non-planes, and estimate their orientation accurately; this is fast and efficient, with application to a real-time system in mind.