Extracting Depth Information From Photographs of Faces

[摘要] Recently new methods of recovering the 3D appearance of objects, like stereo- imaging sensors, laser scanners, and range-imaging sensors provide automatic tools for obtaining the 3D appearance of an object but they require the presence of the object. When only photographic images are available, it is still possible to reconstruct the 3D appearance of the object if there is also a model which can be referenced. The human face is very popular with researchers who try to solve the problems including facial recognition, animation, composition, or modelling. However it is rare to find attempts to reconstruct shape from single photographic images of human faces, although there are numerous methods to solve the shape-from-shading (SFS) problem to date. This thesis describes a novel geometrical approach to reconstructing the original face from a very impoverished facial model1 and a single Lambertian image. This thesis also introduces a different approach to the SFS problem in the sense that it uses prior knowledge of the object, the so-called shape-from-prior-knowledge approach, and addresses the question of what degree of impoverishment is sufficient to compromise the reconstruction. Most recovered surfaces using conventional SFS methods suffer from flattening so that we cannot view them in other directions. We believe that this flatness is due to the lack of geometric knowledge of the subject to be recovered. In this thesis, it is also argued that our approach improves upon existing SFS techniques, because a reconstructed face looks correct even when it is turned to a different orientation from the one in the input image.

[发布机构] University:University of Glasgow

[关键词] Computer science