CAREER: Reconstructing 3D Models from Today?s Scanning Devices

抽象派？3D扫描技术的进步为获取真实的世界中的表面和形状信息提供了一种基本手段，并在从人文学科（如文化保护）到科学（如医学成像）的各个领域发挥了重要作用。这项研究有助于这一趋势，开发算法重建3D模型从原始数据返回今天的采集方式。这项工作中关注的挑战有两个方面：首先，要在实践中有用，算法必须能够处理庞大的数据集，这些数据集通常太大，无法适应应用程序的工作内存。其次，算法必须被设计成通用的，能够适应不同的扫描仪返回的变化和非均匀的数据。为了解决这些挑战，研究人员减少了表面重建的问题，可以解决一个泊松方程，可以解决一个八叉树适合扫描数据。具体来说，这项研究有三个独立的贡献。首先，它描述了一种以局部方式求解全局Poisson系统的新型算法，提供了一种仅需要在任何给定时间在工作内存中维护系统的一小部分集的求解器。其次，它提出了一种新的数据结构，使流遍历通过八叉树，使其能够处理高分辨率的数据，是太大，适合内存。第三，使用有限元公式，它概括了一种模型拟合的方法，该方法允许函数拟合使用非点基元采样的数据，扩展了重建算法支持的采集方式的广度。

AbstractPI ? Kazhdan, MichaelAdvances in 3D scanning technology have provided an essential means for acquiring information about surfaces and shapes in the real world and have played an important role in fields ranging from the humanities (e.g. cultural preservation) to the sciences (e.g. medical imaging). This research contributes to this trend by developing algorithms for reconstructing 3D models from the raw data returned by today's acquisition modalities. The challenges focused on in this work are two-fold: First, to be useful in practice, the algorithms must be able to process huge datasets, often too large to be able to fit into the working memory of an application. Second, the algorithms must be designed to be versatile, capable of adapting to the varying and non-uniform data returned by the different scanners.To address these challenges, the investigators reduce the problem of surface reconstruction to the solution of a Poisson equation which can be solved over an octree adapted to the scanned data. Specifically, this research makes three separate contributions. First, it describes a novel algorithm for solving the global Poisson system in a local manner, providing a solver that only requires a small subset of the system to be maintained in working memory at any given time. Second, it presents a novel data-structure that enables streaming traversal through an octree, making it possible to process high-resolution data that is too large to fit into memory. And third, using a finite elements formulation, it generalizes a method for model fitting that allows functions to be fit to data sampled using non-point primitives, extending the breadth of acquisition modalities supported by the reconstruction algorithm.