题目:城市点云数据理解的对称性和相似性检测
Symmetry and Similarity Detection for Urban Point Cloud Understanding时间:2022年04月29日周五10:00-11:00
地点:腾讯会议 ID:93382698531
腾讯直播间:https://meeting.tencent.com/l/hA1SrLFrwFwl
报告人:薛帆(香港大学,房地产及建设系)
主持人:罗小春(工程管理系)、胡楠(土木工程系)
欢迎广大师生参加!
土木与交通学院
2022年04月25日
报告人简介:
薛帆博士是香港大学房地产及建设系助理教授。薛博士拥有跨学科的教育背景:自动化学士、计算机科学硕士、工业和系统工程博士和建设管理博士后;目前是 ACM会员,CGS高级会员,HKGISA会员,IEEE会员。他的研究兴趣包括:建筑和城市信息建模、激光雷达数据处理、无导数优化、分布式和区块链、以及机器学习。
Dr. Fan (Frank) Xue is an Assistant Professor in the Department of Real Estate and Construction at The University of Hong Kong. Frank has an interdisciplinary background in Automation (BEng), Computer Science (MSc), Industrial and Systems Engineering (PhD), and Construction Management (Postdoctoral fellow – present). He is member of ACM, senior member of CGS, member of HKGISA, and member of IEEE. His research interests include: building and city information modeling, LiDAR data processing, derivative-free optimization, blockchain, and machine learning.
报告摘要:
受益于激光雷达和遥感摄影测量技术的进步,越来越多的建筑和城市拥有了大规模3D 点云数据。 然而,对于提取信息、分割实例和城市计算等应用来说,处理和理解此类非结构化点云数据依然是十分困难的。 我们注意到,许多建筑和城市组件天然地具备了诸如对称性和相似性等几何规律;相形之下,此类由设计与施工原理诱导的非局部特征,比基于卷积的“局部特征”更具特色。 本次演讲将讨论我们最近的面向城市点云理解的对称性和相似性识别的工作。 例如,无论是采用深度学习的“黑盒”形式还是采用无导数优化的“白盒”形式,所识别出的对称性和相似性都能为理解 3D 点数据提供不错的参照信息。
Large-scale 3D point cloud data has become increasingly available for buildings and cities, thanks to the recent LiDAR and photogrammetry technologies in remote sensing. However, processing and understanding such unstructured point data are very challenging for information extraction, instance segmentation, and urban computing. Geometric regularities, such as symmetry and similarity, are naturally equipped by many buildings and city objects, whereas such non-local regularities from design and engineering principles are more characteristic than “local features” in convolution. In this talk, we discuss our recent projects on detecting symmetries and similarities for urban point cloud understanding. Experiments showed that detected symmetries and similarities, whether in “black-box” formulations for deep learning or “white-box” formulations for derivative-free optimization, lead to informative and encouraging results for 3D point data understanding.