Reconstructing façade details using MLS point clouds and Bag-of-Words approach
Thomas Fröch1, Olaf Wysocki1, Ludwig Hoegner1,2, Uwe Stilla1
1Technical University Munich, Germany; 2University of Applied Sciences Munich, Germany
We propose an approach for the reconstruction of 3D fa ̧cade
details. We integrate mobile laser scanning (MLS) point clouds and CAD
models using a Bag of words (BoW) concept, which we augment by
incorporating semi-global features. Our method demonstrates promising
results, improving the conventional BoW approach.
Generating 3D Roof Models from ALS Point Clouds using Roof Line Topologies
Gefei Kong, Hongchao Fan
Norwegian University of Science and Technology, Norway
The automation of 3D roof reconstruction has become a critical research topic in the field of GIScience. Existing methods for this purpose needs to segment roof planes and further extract roof vertices and edges after topology analysis. However, the roof plane-based topology analysis may lead to additional errors for the next step’s extraction result of roof vertices and edges. In this study, based on segmented roof planes, roof edges parallel to the x-y plane are extracted at first, and then the topology relationships of these special roof edges are analyzed and corrected by simple rules. This new approach analyzes the roof structures and extracts roof vertices and edges at the same time, which avoid the accumulated errors by the process of “topology analysis – extraction of roof vertices and edges”. The qualitative and the preliminary quantitative experiment results indicate that the proposed approach can achieve the 3D roof reconstruction well.
MLS2LoD3: Refining low LoDs building models with MLS point clouds to reconstruct semantic LoD3 building models
Olaf Wysocki1, Ludwig Hoegner1,2, Uwe Stilla1
1Photogrammetry and Remote Sensing, TUM School of Engineering and Design, Technical University of Munich, Germany; 2Department of Geoinformatics, University of Applied Science (HM), Munich, Germany
Although LoD3 building models reveal great potential in various applications, they are scarcely available.
In this paper, we introduce a novel refinement strategy enabling LoD3 reconstruction by leveraging the ubiquity of lower LoD building models and the accuracy of MLS point clouds.
Such a strategy promises at-scale LoD3 reconstruction and unlocks LoD3 applications, which we also describe and illustrate in this paper.
Additionally, we present guidelines for reconstructing LoD3 facade elements and their embedding into the CityGML standard model, disseminating gained knowledge to academics and professionals.
Semantic segmentation of buildings using multisource ALS data
Agata Walicka1, Norbert Pfeifer2
1Wrocław University of Environmental and Life Sciences, Institute of Geodesy and Geoinformatics, 50-375 Wrocław, Poland,; 2Department of Geodesy and Geoinformation, Technische Universität Wien, 1040 Vienna, Austria,
Semantic segmentation is a first step of point cloud processing algorithms that are used for many city management applications, such as detection of building footprints, creation of digital twins and city models, preparation of deformation maps and many others. As a result, its accuracy highly influences the results of further processing.
Recently, deep learning approaches for semantic segmentation gained the attention of the community as they enable high classification accuracy with relatively fast processing after the network training phase. However, usually, the network is trained and tested individually for each data set that is processed. Therefore, in this paper, we would like to show a different approach to this problem that includes using two data sets simultaneously for training a deep network. To achieve this goal, we propose to utilize the SparseCNN network and ALS data sets collected for Vienna and Zurich.
The point clouds were classified into ground and water, vegetation, building and bridges, and other classes. The accuracy was tested based on the median IoU value. The results of the experiments showed that including the data from additional source into the training data enabled to keep the high accuracy of vegetation and ground and water classes (94.7% and 97.2%, respectively) while improving the accuracy of buildings and bridges class by around 1 pp and the accuracy of other class by around 1.5 pp (93.3% and 55.1%, respectively).
Classifying point clouds at the facade-level using geometric features and deep learning networks
Yue Tan1, Olaf Wysocki1, Uwe Stilla1, Ludwig Hoegner1,2
1Technical University of Munich, Germany; 2Hochschule München University of Applied Sciences, Germany
Point cloud with classified facade details are key to create digital replicas of the real world. However, few studies have focused on such detailed classification with deep neural networks. We propose a method fusing combining geometric features with deep learning networks for point cloud classification at facade-level. Our experiment concludes that such early-fused features improve deep learning methods' performance.
|
