There are many practical applications where having a three-dimensional (3D) representation of the real world when only 2D data are available will be of great benefit. Unfortunately, the problem of single-view reconstruction is ill-posed and additional constraints are required for its solution. One possible constraint is the so-called 'Manhattan assumption': that the dominant structures of the scene generate linear features that belong to one of three mutually orthogonal directions (e.g., streets, avenues and vertical). In this thesis, I introduce a novel algorithm that uses this assumption to automatically estimate the 3D orientation of a camera relative to the scene. The proposed method relies on a novel Hough transform algorithm that accurately propagates observation uncertainty from image space to parameter space, and uses a novel method for iteratively updating the Hough map as lines are estimated in order to avoid multiple detections of the same scene structure. I evaluate the proposed method on a standard image database and show that it outperforms previous methods.
展开▼
