A study for scalable representation of 3D object models and its applications based on information sensitivity

在这个项目中，我们提出了一种新的3D对象模型表示方法及其应用，该方法基于对3D形状的客观敏感性的概念，可以处理多视角和多分辨率的距离数据、立体图像、纹理图像和视频。客观灵敏度是指当物体的3D形状以一定的细节级别呈现时，可以获得关于该物体的多少信息。该工作包含以下子目标：第一个目标是通过考虑每个测量点的两个确定性来整合在不同分辨率和不同视角下测量的距离数据，然后使用细分和抽取技术等网格自适应技术生成特定分辨率级别的三维形状模型。第二个目标是提出一种统计纹理分析方法，该方法可以从室外场景中拍摄的自然图像中提取一些3D比例因子，而不是像…那样容易无需使用昂贵的工具(如激光测距仪)即可进行3D测量。该方法基于分层线性判别分析，可以对由高阶局部自相关函数计算出的某些特征进行分类。实验证明，该方法能有效地从纹理图像中提取三维尺度因子。我们还构建了一个主动立体视觉系统，可以控制摄像机的平移、倾斜和变焦，作为机器人的智能视觉系统。该系统可以在不需要外界请求的情况下收集本地系统中场景的一些可用信息。第三个目标是开发一种新的三维网格模型之间的交叉参数化技术，可用于各种三维数字几何处理(DGP)。提出了一种基于最小二乘网格技术和自组织可变形模型的交叉参数化方法。这项技术使我们能够将一个3D模型的纹理和运动属性直接转换到另一个模型上，或者在两个3D网格模型之间生成中间模型。作为3D变形在时间上呈现这些中间模型的一种特殊应用，我们做了几个心理学实验，当受试者认识到呈现的中间形状是什么时，他们会回答他们的结果。这些实验表明，他们的认知过程依赖于源和目标对象模型的组合。最后，作为该方法的其他应用，我们开发了基于SDM的人脸增强系统、基于机器学习和聚类算法的3D运动合成系统、基于因式分解方法的3D视频重建系统。较少

In this project, we propose a new method for 3D object model representation and its applications based on a concept of objective sensitivity to a 3D shape, which can deal with multi-view and multi-resolution range data, stereo images, texture images and videos. The objective sensitivity means how much information about the object can be obtained, when its 3D shape is presented in a certain level of detail. The work contains the following sub-goals : The first one is to integrate range data measured in different resolutions and from different view points by considering two certainties for each measured point, and then to generate a 3D shape model in a certain resolution level by using mesh adaptations like subdivision and decimation techniques. Some experimental results have proved that the goal has been achieved.The second goal is to propose a statistical texture analysis method that can extract some 3D scale factor from a natural image taken in an outdoor scene, where it is not so eas … More y to perform 3D measurement without using an expensive tool such as a laser range finder. The method is based on a hierarchical linear discriminant analysis that can classify some features calculated from higher-order local auto-correlation functions. It has been proved that the method is available for extracting 3D scale factor from texture images. We have also constructed an active stereo vision system that can control panning, tilting and zooming of the camera as an intelligent vision system of a robot. This system can gather some available information of a scene in the local system without any request from outside.The third goal is to develop a new cross-parameterization technique between 3D mesh models that can be used in various 3 dimensional Digital Geometry Processings (DGP). The cross-parameterization method proposed is based on a least-square mesh technique and a self-organizing deformable model(SDM) developed by the authors. This technique has enabled us to transfer texture and motion attributes of a 3D model to another one directly, or to generate intermediate models between two 3D mesh models. As a special application of the 3D morphing that presents these intermediate models temporally, we did several psychological experiments where subjects answer their results when they recognize what an intermediate shape presented is. These experiments have shown that their cognitive processes depend on the combination of the source and target object models.Finally, as other applications of the proposed method, we have developed a facial enhancement system based on the SDM, a 3D motion synthesis system based on a machine learning and a clustering algorithm, and a 3D video reconstruction system based on a factorization method. Less