Real-time structure and motion recovery from two views of a multiplanar scene

This demo addresses the problem of recovering the structure of a 3D scene from two views of a scene composed by textured planar surfaces. In this work, we propose a framework that exploits geometrical and algebraic constraints induced by rigidity and planarity in the scene. We present the 3D reconstruction of multi-planar environments as a multi-linear problem expressed in terms of the planar homographies and the projection of intersections between the observed planes. But, instead of solving a complex multi-linear system, our algorithm iteratively solves several linear problems: (i) coplanar feature segmentation, (ii) planar projective transfer estimation, (iii) the epipole computation, and (iv) finding intersections between planes. Linear methods allow our approach to be fast, hence suitable for online 3D reconstruction and real-time localization of a camera moving within a scene.