SBIR Phase I: Reconstructing Consistently Detailed City-Scale Environments From Incomplete 2D and 3D Data

这个小企业创新研究（SBIR）第一阶段项目的更广泛的影响/商业潜力将是使建筑师、城市规划者和房地产开发商（apd）以及许多其他人更便宜、更快捷地使用现实世界的详细模型。APD领域的设计师必须在现实世界的背景下可视化和渲染他们的项目。图片、视频、3D打印甚至虚拟现实为设计过程提供信息，并促进与外行客户和利益相关者的沟通。这些应用程序需要构建世界的一致的详细模型，并且这个项目将自动生成这些模型。我们估计，apd每年至少花费8000万美元手工创建这些模型；而且至少还要花3亿美元在这些模型上，用于模拟、特效和视频游戏设计。通过算法在没有人工干预的情况下生成虚拟模型，将节省手工创建的大量成本（时间和金钱），使设计专业人员可以自由地做他们想做的工作。这个小型企业创新研究（SBIR）第一阶段项目将在为各种商业应用重建高度详细的世界模型方面推进最先进的技术。第一个障碍是解决从嘈杂的点云数据中重建代表现实世界实体（建筑物）边界的表面的问题。虽然表面重建在各种情况下都得到了很好的研究，但它仍然是一个悬而未决的问题，因为成功的算法必须事先了解预期的数据集。使用数据驱动的方法对输入点云进行分割和分类，将有助于对不同对象（例如树木或建筑物）应用不同的重建技术。第二个障碍是机器学习算法的开发，该算法可以处理来自单个统计模型的程序建模和逆程序建模的双重问题，即使无法从源数据中获得该信息（可能在大地理区域内质量不一致），也可以对给定建筑物的细节进行视觉逼真的预测。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will be to make it cheaper and faster for architects, urban planners, and real-estate developers (APDs), as well as many others, to work with detailed models of the real world. Designers in APD fields must visualize and render their projects in the context of the real world. Pictures, videos, 3D printing, and even virtual reality inform the design process and facilitate communication with lay customers and stakeholders. These applications require consistently detailed models of the built world, and this project will automate the generation of these models. We estimate that APDs spend at least $80M annually creating these models by hand; and that at least $300M more is spent on such models for simulations, special effects, and video game design. By algorithmically generating virtual models without human intervention, the significant cost (in time and money) of manual creation will be saved, freeing design professionals to do work they want to be doing.This Small Business Innovation Research (SBIR) Phase I project will advance the state of the art in reconstructing highly detailed models of the world for diverse commercial applications. The first hurdle is solving the problem of reconstructing surfaces representing the boundaries of real-world solids (buildings) from noisy point cloud data. While surface reconstruction is well-studied in a variety of contexts, it remains an open problem in general, as successful algorithms must be informed by priors on the intended datasets. Using a data-driven approach to segment and classify input point clouds will facilitate the application of different reconstruction techniques to different objects (e.g. trees or buildings). The second hurdle is development of a machine learning algorithm which handles the dual problems of procedural modeling and inverse procedural modeling from a single statistical model, enabling visually realistic predictions about the details of a given building, even when that information is not available from source data (which may be of inconsistent quality across a large geographic area).