基于曲面柔韧度的三维形状局部特征描述符研究

In recent decades, with the development of capturing, modeling, and visualizing techniques for 3D shapes, 3D model has been widely used in various application fields especially on the Internet and in domain-specific databases. It is important to develop 3D shape retrieval system which retrieves similar 3D object by a given querying object. 3D shape retrieval is not an easy problem because 3D shapes have less descriptive information especially in the case of cluttered scenes, or when dealing with partial objects, or skeletal articulation changes with deformations. Many descriptors and matching methods have been developed for overcoming the problems. Comparing with global matching, although partial shape matching has more technical challenges, it can overcome limitations of global matching. Furthermore, many applications need to match or retrieval partial 3D shape with non-rigid transformations. Therefore, partial shape matching with non-rigid transformation is an interest research field.. Shape descriptors are of paramount importance in partial shape matching, and they are commonly used to tackle these problems. The most used feature parameters is curvature, but it has a critical limitation that it extremely sensitive to small perturbations in the surface smoothness and to quantization effects when they operate on triangulated surfaces. Thus, in this research project we propose a multi-resolution approach called surface flexibility, which not only estimates the curvature of a vertex over neighborhoods of variable sizes, but also takes into account the topology of the surface in the neighborhood. Our approach is based on calculating ratio between the volume of intersection between sphere and the model, and the volume of sphere. In addition, we will employ the multi-scale surface flexibility parameter to estimate the saliency of mesh for reducing the computation complexity of matching 3D objects. We also consider to research a generative model to encode several local descriptors to a more discriminative local descriptor. In this research, we consider to utilize surface flexibility, shape diameter function, and heat kernel signature to compose a generative local shape descriptor. Because each local descriptor have its advantages, and complementary effect can be achieved, better shape descriptive capability and matching accuracy can be accomplished.. We will adopt TOSCA data-set and SHREC 2011 benchmark to evaluate the performance of researched methods, and also compare with some representative methods in this research field. The proposed methods provide a fundamental theory and tools to three-dimensional shape analysis fields, which including matching, retrieval, corresponding, segmentation, symmetry analysis, skeletonization, and so on. It will bring advantages to the development of related research fields.

三维形状局部特征描述符是图形学与模式识别领域的一项重要研究课题。针对目前既存的局部特征描述符所面临的区分度不高、受限于形状非刚性变形等问题，本研究首先拟以三维形状的基本特征参数为基础，提出一种基于曲面柔韧度的局部特征描述，以克服以往方法对局部噪声敏感的缺陷，并增加多尺度特征描述能力，能很好的应对物体刚性和非刚性变形，为三维形状分析提供基础；研究基于多尺度曲面柔韧度的局部显著区域提取方法，提高形状显著区域的提取能力，并降低部分形状匹配与检索过程的计算复杂度；此外，研究一种多类型局部特征描述符的整合方法，利用隐马尔科夫模型训练模型参数，综合提升局部特征描述能力和鲁棒性。本研究计划采用TOSCA和SHREC11形状数据库进行性能评价实验，与代表性方法进行比较并分析本研究方法的优缺点。该研究成果可以应用到三维形状部分匹配、检索、对应、分割等领域，对推动相关领域的发展起到积极的作用。

人类主要通过双眼来观察世界，通过三维的方式分析所处位置的环境、周围物体等，从而实现高度的智能。为实现和人类似的空间感知和智能，机器学习和计算机视觉是目前主要的解决方法，三维形状局部特征描述符是图形学与模式识别领域的一项重要研究课题。本课题针对三维物体的识别、认识、感知开展研究工作，通过分析目前既存的局部特征描述符所面临的区分度不高、受限于形状非刚性变形等问题，研究区分度、描述能力更高的征描述符，从而提升三维形状的识别精度，为实现更高精度的场景解析、物体识别等提供技术手段。主要的研究内容包括：1）三维形状局部特征描述符提取，研究基于曲面柔韧度的特征表达，在此基础上研究基于深度学习的特征提取方法；2）显著区域的识别与检测，研究形状显著性的表达方式，构建机器学习模型；3）多种特征表达的自适应融合，基于多模态深度学习的三维形状特征融合；4）实验与性能评估。通过三年不懈的努力，项目取得了一定的成果，达到了预期的目标。研究过程中提出了一些原创的研究方法，包括3D Shape Deep Learning, Multi-modal Deep Learning for 3D Shape, Shift-Invariant Ring Feature等方法，项目开展过程中发表了16篇科研论文，申请了10个专利，其中14篇论文被SCI收录。关于三维模型的特征提取，我们发表在IEEE TMM上的论文，较早地将深度学习成功应用到三维形状特征提取，解决了三维形状无法使用深度学习这一个国际难题。通过在标准数据集上测试，研究方法取得了较好的识别精度；参加了Eurographics Workshop for 3D Object Retrieval SHREC 2014比赛，在国际众多研究方法中，取得了较好的成绩。通过此项目的研究，探明了三维形状的特征的深度学习，解决了三维形状的非均一拓扑连接关系的深度学习；由于使用深度学习，能够提高特征的表述能力，能够克服手工设计特征的性能不稳定、表征能力弱等问题；研究的多模态深度学习，更进一步揭示了视觉对三维物体分析的机理。研究成果对三维形状分析领域有一定的支撑作用，一定程度上惠及几乎所有依赖视觉信息理解的应用。此外本项目提出的三维深度学习的新思路，能够进一步扩展深度学习理论和应用范围。