Wang, RuishengWang, XinYang, Hongxin2024-10-072024-10-072024-10-04Yang, H. (2024). Three-dimensional building reconstruction from ALS point clouds (Doctoral thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.https://hdl.handle.net/1880/119951Reconstructing buildings from Light Detection and Ranging (LiDAR) point clouds obtained from aerial perspectives is of significant importance in the domain of photogrammetry. Given that the experimental dataset, Building3D, lacks sufficient corner points and exhibits point cloud sparsity among other challenges, point cloud completion (PCC) techniques, a branch of reconstruction, are employed to complete the building facade information. Due to the high demand for labeled data and the associated high cost of manual annotations, Self- Supervised Learning (SSL) methods for three-dimensional (3D) point clouds have garnered considerable attention from scholars. However, existing methods commonly use a standard Transformer backbone, result- ing in quadratic time complexity. To overcome these limitations, an innovative masked linear autoencoder framework is proposed. Due to the storage requirements—approximately 400:4:1 for point cloud, mesh, and wireframe formats, respectively, wireframe models have recently garnered considerable attention in the field of remote sensing. Despite some early explorations into constructing wireframe models, numerous challenges persist. This thesis revisits 3D building wireframe reconstruction from a SSL perspective, with the aim of alleviating or even addressing these existing difficulties. A two-stage Self-supervised (SS) pretraining architecture is proposed to generate wireframe models. Initially, it utilizes a SSL-based pretraining framework that incorporates a linear self-attention mechanism (SAM) to generate point-wise features. Subsequently, corner detection and edge prediction modules are employed to classify and regress the coordinates of corner points and to determine optimal edge selections, respectively. To address the issue of insufficient corner points, a SSL-based pretraining method for 3D wireframe reconstruction, guided by an edge point regression module, is proposed. The parameters of the wireframe’s edges—including edge length, direction vector, and direction offset—are regressed under the guidance of the edge point regression module. To enhance the clustering of roof wireframe vertices, an efficient approach based on a multiclass TWin Support Vector Machine (TWSVM) framework is proposed. This framework aims to simplify the model by effectively clustering roof wireframe vertices.enUniversity of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.Building reconstructionpoint cloudsRemote Sensingdoctoral thesis