Generic and Real-Time Structure from Motion

Abstract

We introduce a generic and incremental Structure from Motion method. By generic, we mean that the proposed method is independent of any specific camera model. During the incremental 3D reconstruction, parameters of 3D points and camera poses are refined simultaneously by a generic local bundle adjustment that minimizes an angular error between rays. This method has three main advantages: it is generic, fast and accurate. The proposed method is evaluated by experiments on real data with three kinds of calibrated cameras: stereo rig, perspective and catadioptric cameras.