3D reconstruction from uncalibrated images taken from widely separated views

In this paper, we present a framework for 3D reconstruction based on uncalibrated images taken from widely separated views. Our method starts from scale-invariant key points being detected and described, then several schemas to improve the key points matching results being adopted. Consequently, with the fundamental matrix estimated from the key point correspondences, the epipolar geometry constraints between each view are recovered. We refine correspondence result by epipolar line and affine-invariant constraints. As a result, the refined correspondences will improve the fundamental matrix estimation. With the recovered fundamental matrix and epipolar, the sparse projective 3D point cloud of the scene could be recovered. After that, a globally nonlinear optimal procedure combined with Interval Analysis technique is performed to upgrade the projective 3D points to metric structure. The experimental results show our framework is effective for 3D reconstruction task.