Supervisors: Alex Locher, Dr. Ajad Chhatkuli, Prof. Luc Van Gool
We present a new approach for single image 3D surface reconstruction in the context of document scanning, where a complete 3D geometric prior is not available. The state-of-the-art methods in this domain make use of a 3D template, incorporating the complete geometry and texture of the surface in its pre-deformation state. In many practical cases, a complete 3D template is not known and/or the feature point matching between the 3D template and the deformed surface may be unreliable. This makes the problem of single image 3D reconstruction a highly challenging one. In this thesis, we study how minimal geometric knowledge like the aspect ratio of a document can be combined with several visual cues and geometric deformation priors to accurately reconstruct the surface from a single input image. We use sheets of paper as well as store receipts in various states of deformation as the objects to be reconstructed. This gives this thesis a promising practical application, where the 3D reconstruction can be used to obtain a high quality flat template. We give practical methods on exploiting the deformation priors based only on projected contour matching. Furthermore, an approach is presented where we efficiently combine unreliable visual cues such as shading and texture with the isometric deformation prior, to obtain reconstruction accuracy unattainable by the state-of-the-art methods.