Probabilistic scanning for high accuracy hand-held 3D reconstruction

这项拨款提案属于计算机视觉领域。它的主要重点是开发新的主动式3D重建技术，这些技术对材料属性的变化具有高精度、快速和健壮的性能，例如互反射(光泽材料)和亚表面散射(半透明材料)。3D主动扫描已经存在很长一段时间了，但由于城市的3D地图、游戏控制(Kinect)和3D打印日益普及等新应用，它最近获得了新的动力。其目标是创建新的结构光算法，适用于手持便携式扫描仪，其便携性不会影响准确性和健壮性。手持式扫描技术的应用范围很广，从工业逆向工程到质量控制、损害评估或医疗用途的形状估计。基本算法的进步将与应用程序开发密切相关，例如为技术艺术中的自适应和建筑投影建立场景模型，以及扫描3D打印。随着3D打印的大众化，出现了对实用的逆向工程解决方案的需求，以允许高效和准确地扫描各种对象。此外，自动打印质量控制和打印机校准目前很难实现，因为3D打印机创建的塑料物体本来很难扫描，因为它们使用的是有光泽或半透明的材料。为了充分发展，3D打印领域将需要非常准确、健壮、快速、灵活和低成本的方法来扫描打印对象(如果可能的话，在打印过程中)以及真实对象。目前的扫描仪技术不能达到所有这些要求，所以这就是这笔赠款打算通过探索新的方法来进一步推进3D扫描领域的贡献。显然，执行3D扫描的方法有很多种，但我们认为最有用和最有前景的方向是实时手持扫描仪。为了换取其提供的灵活性、便携性和效率，手持扫描仪的精确度和健壮性都有所下降。这主要是因为这些扫描仪不仅需要为3D三角测量建立匹配，而且在扫描进行时还需要准确地跟踪自己在太空中的位置。这项研究将从我们最新的结果出发，其中提出了一种新的概率方法，用于使用新形式的结构光建立密集对应。本研究涉及的子课题包括空间相干分析、概率高维匹配算法、增量式概率扫描、多摄像机系统和建筑多投影仪系统。

This grant proposal is in the field of computer vision. Its primary focus is on developing new active 3D reconstruction technologies which are highly accurate, fast, and robust to variations in material properties such as inter-reflection (glossy materials) and sub-surface scattering (translucent materials). 3D active scanning has been around for a long time, but it has recently gained new momentum because of new applications such as 3D mapping of cities, gaming controls (Kinect), and the increasing accessibility of 3D printing. The goal is to create new structured light algorithms suitable for a hand-held portable scanner, where portability will not compromise accuracy and robustness. Applications of hand-held scanning technology are numerous and range from industrial reverse engineering to quality control, damage assessment, or shape estimation for medical use. Fundamental algorithmic advances will be closely matched by applications development, such as scene modelling for adaptive and architectural projections in technological arts, and scanning for 3D printing. With the democratization of 3D printing, a need is emerging for practical reverse engineering solutions to allow scanning of various objects efficiently and accurately. Also, automated print quality control and printer calibration are currently difficult to achieve, since 3D printers create plastic objects which are inherently hard to scan because they use glossy or translucent materials. In order to develop fully, the field of 3D printing will require very accurate, robust, fast, flexible and low cost methods for scanning printed objects (during the printing process, if possible) as well as real objects. Current scanner technologies can not achieve all of these requirements, so this is where this grant intends to contribute by exploring new approaches to further advance the field of 3D scanning. Obviously, there are many ways to perform 3D scanning, but we consider that the most useful and promising direction is toward real-time hand-held scanners. In exchange for the flexibility, portability and efficiency they offer, hand-held scanners suffer from reduced accuracy and robustness. This is mostly due to the fact that these scanners need not only to establish matches for 3D triangulation, but also accurately track their own position in space as the scan proceeds. This research will proceed from our most recent results where a new probabilistic approach was proposed for establishing dense correspondence using a new form of structured light. The various sub-topics involved in this research are spatial coherence analysis, probabilistic high-dimension matching algorithms, incremental probabilistic scanning, multi-camera systems, and architectural multi-projector systems.