We present Strokes2Surface, an offline geometry-reconstruction pipeline built upon a 4D Sketching Interface, MR.Sketch, targeted at architectural design. The pipeline recovers a curve network from designer-drawn strokes, thus bridging between concept design and digital modeling stages in architectural design. The input to our pipeline consists of 3D strokes' polyline vertices and their corresponding timestamps (as of the fourth dimension), along with additional geometric and stylus-related recorded properties. Inspired by sketch consolidation and sketch-based modeling methods, our pipeline leverages such data and combines three Machine Learning (ML) models; a classifier and two clustering models. In particular, based on observations of practices designers typically employ in architectural design sketches, we solve a binary classification problem to recognize whether a stroke depicts a boundary and edge or is used to fill in the enclosing areas and faces of the intended architectural object. Followed by the two clustering models, strokes of each type are further parsed into groups, each representing either a single edge or a single face. Next, groups representing edges are approximated with B-spline curves, followed by a topology-recovering process identifying and fixing desired connectivities between the curves forming a well-connected curve network. Next, groups representing the faces are employed to detect the cycles bounding patches in the curve network, resulting in the final surface mesh geometry of the architectural object. We confirm the usability of Strokes2Surface via a user study and further validate and compare our results against a range of reconstructions computed using alternative methods. We also introduce our manually labeled dataset of 4D architectural design sketches for further use in the community.

翻译：我们提出Strokes2Surface，这是一种基于4D草图界面MR.Sketch构建的离线几何重建管线，专为建筑设计领域设计。该管线从设计师绘制的笔画中恢复曲线网络，从而弥合建筑设计中的概念设计与数字建模阶段之间的鸿沟。管线的输入包括3D笔画的多段线顶点及其对应时间戳（作为第四维度），以及额外的几何和触控笔相关记录属性。受草图整合与基于草图的建模方法启发，我们的管线利用此类数据，结合三种机器学习（ML）模型：一个分类器和两个聚类模型。具体而言，基于对建筑师在建筑设计草图中常用实践的观察，我们解决一个二分类问题，以识别笔画是在描绘边界和边缘，还是用于填充目标建筑对象的封闭区域与面。随后，通过两个聚类模型，将每种类型的笔画进一步解析为组，每组分别代表单条边缘或单个面。接着，代表边缘的组以B样条曲线进行近似拟合，再通过拓扑恢复过程识别并修复曲线之间所需的连接关系，形成良好连接的曲线网络。最后，利用代表面的组检测曲线网络中的面片边界环路，生成建筑对象的最终曲面网格几何体。我们通过用户研究验证了Strokes2Surface的可用性，并进一步基于使用多种替代方法计算的重建结果进行对比验证。我们还引入了一个手动标注的4D建筑设计草图数据集，供社区进一步使用。