摘要
建筑立面解析对于城市分析、语义重建、能源需求评估等需要高质量语义数据的任务至关重要,而传统的手动测量及建模方式往往十分耗时.本文探讨了基于迁移学习的Mask R-CNN网络对建筑外立面图像中的门窗进行自动坐标提取的方法.数据集方面选用当前基准数据集中标注图片数量最多的CMP数据集并为适应训练网络对原数据集进行了数据格式转换.结果表明该方法可在较高精准度基础上实现对建筑立面图像中门窗角点坐标的格式化输出.通过这项工作,希望为倾斜摄影等高细节层次的建筑三维建模领域的研究提供解决思路.
Abstract
Parsing building facades is crucial for tasks such as urban analysis,semantic reconstruction,and energy demand assessment,which require high-quality semantic data.However,traditional manual measurement and modeling methods are often time-consuming.This article explores a method based on transfer learning with Mask R-CNN networks for automatic extraction of door and window coordinates in building facade images.The dataset selected for this study is the CMP dataset,which is currently the benchmark dataset with the highest number of annotated images,and it was adapted to train the network by converting the data format.The results demonstrate that this method can achieve formatted output of door and window corner coordinates in building facade images with high precision.Through this work,we aim to provide a solution approach for research in the field of high-detail architectural 3D modeling,such as oblique photography.
NETL
NSTL
万方数据