Single-Image Based 3D Modeling Techniques And Applications

3D modeling is an important part of computer graphics. Traditional 3D modeling techniques are mainly divided into two categories. The first category requires profession-al users (such as artists or designers) and softwares (such as 3DMax), the other category requires special devices (such as 3D laser scanners) to obtain multiple images or 3D in-formation of the scene. These techniques are not suitable for ordinary users. This thesis presents new modeling techniques that, instead of special devices, softwares or complex user interaction, only require a single image as input. Hence, to a large extent, the use for ordinary users is facilitated and the application of modeling algorithms is extended.The main difficulty of constructing 3D models from a single image is the inadequa-cy of the input information, for example, the input image contains only color or depth information at a single view. Building a complete 3D model from such inadequate input information usually requires a lot of user interaction. The research work carried out in this thesis greatly reduces or even avoids user interaction in the modeling process by ex-ploiting prior knowledge from specific to common domains. The work consists mainly of three parts:· We study 3D modeling with the input of single vectorgraph or sketch. Taking the sketch of furniture assembly instructions as an example, a set of novel vectorgraph processing and modeling algorithms is proposed in this part. The proposed algo-rithms combine geometric and semantic priors of 3D furniture components, and successfully construct functional 3D furniture from the input. Moreover, we com-bine sketches of multiple assembly steps to extract the dynamic assembly process,resulting in a complete 3D assembly animation, which effectively assists the user's own furniture assembly.· We study 3D shape completion from a single RGBD image. Using voxels to ex-press 3D objects and combining two priors of common 3D objects, global simi-larity of the same category and local similarity of the same object,we propose a novel voxel-match based global optimization algorithm that, starting from a sin-gle RGBD image of any category, automatically constructs a complete 3D object with geometric details. The verification results on standard data sets and other real-world data illustrate the effectiveness and robustness of the algorithm.· We study the modeling of 3D lines from a single RGB image, which are used to improve 2D image warping techniques. We first use projective geometry and vision theories to extract visually sensitive 3D lines from the input 2D image, then we combine the 3D information into traditional 2D warping energy functions and propose an efficient algorithm to solve the new energy function. The new warping algorithm effectively preserves the overall information of the 3D lines and reduces unnatural distortion of the warping results. The usefulness and effectiveness of the algorithm are validated in a variety of domains such as panoramas, wide-angle and perspective images, and part of general images.