Efficient 3D object representation and reconstruction using depth/texture maps

Abstract

This dissertation introduces an effective representation method of 3D shape and texture which will be called DoSurface/ToSurface (Depth/Texture on Surface) and a depth estimation algorithm which is appropriate to the representation in image-based rendering environment using multi-view input images. On this framework, we will show that this representation is also useful for object reconstruction. While conventional reconstruction methods focus on the estimation of 3D information only, we consider compactness and effectiveness of the information, and reconstruction results as well. DoSurface/ToSurface represents 3D information on intermediate surface by multiple-projection to that surface. It is also a good directing post for mesh generation of 3D reconstruction. We first suggest DoSurface/ToSurface representation of 3D information. Although a projection method of 3D information to cylinder surface has been suggested, it has a disadvantage that inner surface of an object is not represented. In DoSurface/ ToSurface representation, depth values in pre-defined positions are obtained, unnecessary information is removed and surface of 3D object is represented without any loss of information. Secondly, a depth estimation algorithm, named as Projected Line Search (PLS) algorithm, suitable for DoSurface is suggested as well. PLS is faster and more accurate than other space carving algorithms because it directly computes the surface position of an object. Thirdly, a fast algorithm of filtering for DoSurface is suggested. The proposed method reduces a considerable amount of execution time and it requires only the small memory. Fourthly, DoSurface/ToSurface is applied to reconstruction process, and we search a mesh generation method suitable for this representation. It will be shown that the representation is also useful for reconstruction.