Representation and reconstruction of 3D shapes in computer vision

Given a single image taken from any popular digital camera, how can an object from the scene be reconstructed, or "grabbed", from the picture as a three-dimensional (3D) object? We humans do it effortlessly! This thesis explores how to give computers such a 3D perception. We study how to efficiently represent 3D objects to allow them to be reconstructed from an image. This is challenging as there might exist different shapes that have similar 2D projection. To constrain this ill-posed problem, we propose novel methods to encode the object's geometry and to reconstruct them by relying on 3D prior information.