Abstract
Due to the rapid cost down of 3D display devices, 3D multimedia gradually comes into our daily life in recent years. However, the promotion of 3D multimedia and related applications are still restricted by insufficient stereoscopic contents. By contrast, we notice that 2D materials are richer and easily available, therefore, we can fairly conclude that generating 3D contents from these existing materials is a good idea. To accomplish our purpose, it is necessary to design a mathematical model to describe the relation between an image and its corresponding scene. In the past 10 years, many researches about 2D-to-3D conversion have been published such as depth-from-texture, depth-from-focus, depth-from-shading and depth-from-motion, but most of them consider only few of depth cues. According to the work done by Sexena et al., we use machine learning to combine a lot of depth cues as our feature vector, and explore scene structure with high level computer vision, e.g. natural boundary detection and surface classification. In detail, the basic unit of a scene forms a small plane called ‘superpixel’ that aggregates similar pixels. All depth cues and scene information, which correspond to local term and smoothness term respectively, are combined using a high order Markov random field. Because we chose L2-norm as our error function, the problem to find optimal depthmap could be solved using quadratic programming. Consequently, we proposed a 2D-to-3D conversion algorithm considering high level and low level vision cues for generation of 3D contents automatically. We can regard it as a tool to promote 3D multimedia and advanced image analysis.