CHS: Small: Shape Processing with Deep Architectures

三维形状的数字表示正成为许多科学和工程领域不可或缺的一部分。3D打印机正变得越来越流行，可以将这些数字表示转换为工业或家庭使用的实物，作为工具、机械部件或家庭用品。用于模拟、教育和娱乐的虚拟环境需要大量不同的3D模型来保持真实感和可操作性。在计算机视觉中，最近的几种对象识别算法是在来自3D形状存储库的大型合成数据集上进行训练的。所有这些领域以及其他几个领域越来越依赖于数字3D图形的可用性。然而，开发允许用户轻松生成新的3D形状，甚至从在线存储库中检索和处理现有形状的工具仍然是一个挑战。当前的3D建模工具通常需要通过低级选择和编辑命令进行费力的用户交互，而现有的用于检索3D形状的搜索引擎在很大程度上依赖于人工输入的标签和手动调整的特征表示，这导致检索性能不令人满意。PI在这项研究中的目标是创建基于“深度”架构的算法，该算法自动从数据中学习如何可靠地分析和合成针对3D形状处理和合成性能进行优化的3D形状(因此，例如，当这些形状被3D打印时，它们将保留所需的物理特性，如降低机械应力)。项目成果将以开源形式发布，不仅将对计算机图形学和计算机视觉产生广泛影响，而且将对工业、计算机辅助设计和机械工程管道以及建筑和室内环境的建筑工程软件产生广泛影响。这项研究将在智力方面做出三大贡献：1)深层体系结构和学习三维形状特征表示的算法优化了检索和处理性能；算法将在海量2D图像数据集和3D模型库上进行训练，学习的表示将用于准确的形状分类、分割、对应、样式分析和纹理。2)深层架构和算法，用于学习基于直观的用户输入(例如草图和文本描述)可靠地生成新的3D形状。3)用于学习优化3D形状的底层几何以使其获得3D打印和制造所需的物理属性的算法。

Digital representations of three-dimensional shapes are becoming an integral part of many scientific and engineering fields. And 3D printers are becoming increasingly popular for transforming these digital representations into real objects for industrial or home use as tools, mechanical components, or household items. Virtual environments for simulation, education, and entertainment require large numbers of diverse 3D models to maintain realism and operability. In computer vision, several recent object recognition algorithms are trained on large synthetic datasets originating from 3D shape repositories. All these fields and several others increasingly depend on the availability of digital 3D shapes. However, it is still a challenge to develop tools that allow users to easily produce new 3D shapes, or even to retrieve and process existing ones from online repositories. Current 3D modeling tools often require laborious user interaction via low-level selection and editing commands, while existing search engines for retrieving 3D shapes largely depend on manually entered tags and hand-tuned feature representations which results in unsatisfactory retrieval performance. The PI's goal in this research is to create algorithms based on "deep" architectures that automatically learn from data how to reliably analyze and synthesize 3D shapes that are optimized for 3D shape processing and synthesis performance (so that, for example, when these shapes are 3D printed they retain desirable physical properties such as reduced mechanical stress). Project outcomes will be released as open source code and will have broad impact not only on computer graphics and computer vision but also on industry, computer-aided design and mechanical engineering pipelines, and architectural engineering software for buildings and indoor environments.This research will make three major contributions in terms of intellectual merit: 1) Deep architectures and algorithms for learning 3D shape feature representations optimized for retrieval and processing performance; the algorithms will be trained on massive 2D image datasets as well as 3D model repositories, and the learned representations will be used for accurate shape categorization, segmentation, correspondence, style analysis, and texturing. 2) Deep architectures and algorithms for learning to reliably generate new 3D shapes based on intuitive user input, such as sketches and textual descriptions. 3) Algorithms for learning to optimize the underlying geometry of 3D shapes such that they acquire desired physical properties for 3D printing and manufacturing.