Structure from Motion with Higher-level Environment Representations

Abstract

Computer vision is an important area focusing on understanding, extracting and using the information from vision-based sensor. It has many applications such as vision-based 3D reconstruction, simultaneous localization and mapping(SLAM) and data-driven understanding of the real world. Vision is a fundamental sensing modality in many different fields of application.

While the traditional structure from motion mostly uses sparse point-based feature, this thesis aims to explore the possibility of using higher order feature representation. It starts with a joint work which uses straight line for feature representation and performs bundle adjustment with straight line parameterization. Then, we further try an even higher order representation where we use Bezier spline for parameterization. We start with a simple case where all contours are lying on the plane and uses Bezier splines to parametrize the curves in the background and optimize on both camera position and Bezier splines. For application, we present a complete end-to-end pipeline which produces meaningful dense 3D models from natural data of a 3D object: the target object is placed on a structured but unknown planar background that is modeled with splines. The data is captured using only a hand-held monocular camera. However, this application is limited to a planar scenario and we manage to push the parameterizations into real 3D. Following the potential of this idea, we introduce a more flexible higher-order extension of points that provide a general model for structural edges in the environment, no matter if straight or curved. Our model relies on linked B´ezier curves, the geometric intuition of which proves great benefits during parameter initialization and regularization. We present the first fully automatic pipeline that is able to generate spline-based representations without any human supervision. Besides a full graphical formulation of the problem, we introduce both geometric and photometric cues as well as higher-level concepts such overall curve visibility and viewing angle restrictions to automatically manage the correspondences in the graph. Results prove that curve-based structure from motion with splines is able to outperform state-of-the-art sparse feature-based methods, as well as to model curved edges in the environment.