Reliable 3D reconstruction from a few catadioptric images

This paper proposes a scheme for the reliable reconstruction of indoor scenes from a few catadioptric images. A set of hand-detected correspondences are established across (but not necessarily all) images. Our improved method is used for the estimation of the essential matrix from appropriately normalized point coordinates. Motion parameters are computed by using the Hartley's (1993) decomposition. A heuristic method is suggested for selecting the point pairs which are most reliable for 3D reconstruction. The known mid-point method is applied for computing the 3D model of a real scene. The parameters of the catadioptric sensor are approximately known but no precise self-calibration method is performed. The experiments show that a reliable 3D reconstruction is possible even without complicated non-linear self-calibration and/or reconstruction methods.