Structure from Motion

Abstract

This chapter presents the computational framework of 3D reconstruction from 2D image points or structure from motion. We first present the standard sparse approach, which is sufficient for computing camera poses or camera tracking, but insufficient for broad modeling and visualization applications that require a higher density of reconstructed points. We then introduce a resampling scheme to obtain point correspondences at sub-pixel level, and develop a unique quasi-dense approach to structure from motion. The quasi-dense approach results in a more robust and accurate geometry estimation using fewer images than the sparse approach by delivering a higher density of 3D points. The modeling applications presented in Part III of the book are based on this quasi-dense reconstruction.