基于单幅图像的三维建模技术及其应用

发布时间：2018-05-15 12:27

本文选题：三维建模 + 单幅图像　；参考：《浙江大学》2017年博士论文

【摘要】：三维建模是计算机图形学研究的核心内容之一。传统的三维建模技术主要分为两类,一类需要由专业用户(如美工或设计师)使用专业软件(如3DMax)来建模,另一类则需要使用专业设备(如三维激光扫描仪)来获取场景的多幅图像或三维信息,在此基础之上构建三维模型。这些技术不适用于普通用户使用。本文研究了从单幅图像输入进行三维建模的技术,不需要专业软件、专业设备以及复杂的用户操作,极大地方便了普通用户使用这些技术,推广了建模算法的应用范围。从单幅图像中构建三维模型的主要困难在于输入信息的不充分性,例如输入图像中只包含了单一视角下的颜色或深度信息。从这样不充分的输入信息中构建出完整的三维模型通常需要大量的用户交互。本文开展的研究通过利用从特定领域到通用领域的先验知识,大大减少甚至彻底避免了建模过程中的用户交互。这些工作主要由三个部分组成:·研究了以单幅矢量草图为输入的三维建模技术。我们以家具组装说明书的草图为例,提出了一套新的矢量草图处理和建模技术。所提出的算法结合三维家具部件的几何形状先验和部件之间的语义相关性先验,从单幅矢量草图出发成功地建模了功能正确的三维家具。更进一步的,我们结合组装说明书不同组装步骤的多页草图,提取了家具组装的动态过程,生成了完整的三维组装动画,有效辅助了普通用户自己进行家具组装。·研究了以单幅RGBD图像为输入的三维形状补全技术。我们利用体素来表达三维物体,结合同一种类物体形状的整体相似性和同一物体局部几何细节的自相似性这两个先验知识,提出了一个新的基于体素块匹配的全局优化算法,可以全自动地从任意种类物体的单一视角图像出发得到一个完整的、具有几何细节的三维模型。在标准测试数据集和其他拍摄数据上的验证结果说明了我们算法的有效性和鲁棒性。·研究了从单幅RGB图像建模三维线条并利用这些三维线条辅助二维图像变形的技术。我们首先利用投影几何原理和视觉技术提取二维输入图像中的视觉敏感的三维线条信息,然后加入到传统的二维变形的能量方程中并给出了新的高效的求解算法。新的变形算法有效保持了三维线条的整体信息,减少了图像变形结果的不自然扭曲。算法的实用性和有效性在全景图、广角图、强透视图以及部分一般图像的多种应用中得到验证。
[Abstract]:Three-dimensional modeling is one of the core contents of computer graphics research. Traditional 3D modeling techniques are mainly divided into two categories. One requires professional users (such as art workers or designers) to use professional software (such as 3D Max) to model. On the other hand, professional devices (such as 3D laser scanner) are used to obtain multiple images or 3D information of the scene, and the 3D model is constructed on the basis of this. These technologies are not suitable for ordinary users. In this paper, the technology of 3D modeling from single image input is studied. It does not require professional software, professional equipment and complex user operation, which greatly facilitates the ordinary users to use these techniques, and extends the application of modeling algorithm. The main difficulty of constructing 3D model from a single image is the inadequacies of input information, for example, the input image contains only the color or depth information under a single angle of view. It usually takes a lot of user interaction to build a complete 3D model from such inadequate input information. By utilizing the prior knowledge from a specific domain to a general domain, the research in this paper greatly reduces or even avoids the user interaction in the modeling process. These works are mainly composed of three parts: a 3D modeling technique with single vector sketch as input is studied. Taking the sketch of furniture assembly manual as an example, a new set of vector sketch processing and modeling techniques is proposed. Based on the prior geometric shape of the 3D furniture components and the prior semantic correlation between the components, the proposed algorithm successfully models the three dimensional furniture with correct function from a single vector sketch. Furthermore, we combined the multi-page sketches of different assembly steps of the assembly instructions, extracted the dynamic process of the furniture assembly, and generated a complete three-dimensional assembly animation. It effectively assists ordinary users to assemble furniture by themselves, and studies the 3D shape completion technique with single RGBD image as input. In this paper, we propose a new global optimization algorithm based on volume block matching, which uses volume primer to represent 3D objects, combining the two prior knowledge of the global similarity of the shape of the same object and the self-similarity of the local geometric details of the same object. A complete 3D model with geometric details can be obtained automatically from a single view image of any kind of object. The validation results on the standard test data set and other photographed data show the effectiveness and robustness of our algorithm. The technique of modeling 3D lines from a single RGB image and using these lines to assist 2D image deformation is studied. Firstly, we use the principle of projection geometry and visual technology to extract the vision-sensitive 3D line information from the two-dimensional input image, then we add it to the traditional energy equation of two-dimensional deformation and give a new efficient algorithm to solve the problem. The new algorithm effectively preserves the whole information of 3D lines and reduces the unnatural distortion of image deformation results. The practicability and validity of the algorithm are verified in many applications of panorama, wide angle, strong perspective and some general images.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】