基于几何先验的单图像交互式分段平面建筑重建

从单张图像重建建筑物的分片平面结构是计算机视觉领域中一个重要且具有挑战性的问题。尽管许多研究都致力于完全自动化的重建，但在复杂的建筑结构上的结果往往不尽人意。为了解决这一问题，我们提出了一种交互式的单图像3D建筑重建方法，该方法结合了基于几何先验的自动平面推断和交互式平面细化。利用卷积神经网络，可以推导出如平面方向和平面之间的交叉角度等几何先验知识。基于这些知识，通过点击操作，建筑区域首先被交互式地划分为带有结构信息的多个多边形区域。然后，通过充分利用几何先验所提供的约束，自动并依次推断与这些多边形部分相关的平面。最终，在点击交互的指导下，利用图像线索和几何先验，导出的平面进行全局优化。在三个数据集上的实验结果证实，我们的方法在定性和定量方面都超过了现有的技术。

Reconstructing the piecewise planar structure of buildings from a single image is an important and challenging problem in the field of computer vision. Although numerous studies have focused on fully automated reconstruction, the results on complex architectural structures are often unsatisfactory. To address this issue, we propose an interactive 3D building reconstruction method from a single image, which combines automated plane inference based on geometric priors and interactive plane refinement. By leveraging convolutional neural networks, geometric prior knowledge such as plane orientations and the crossing angles between planes can be derived. Based on this knowledge, through click interactions, building regions are first interactively partitioned into multiple polygonal regions with structural information. Subsequently, by fully utilizing the constraints provided by the geometric priors, planes associated with these polygonal segments are automatically and sequentially inferred. Finally, guided by the click interactions, the derived planes undergo global optimization by leveraging image cues and geometric priors. Experimental results on three datasets confirm that our method outperforms existing state-of-the-art techniques in both qualitative and quantitative aspects.