With the LiDAR developments, point cloud is becoming a valuable resource to build 3D models of Digital Twins, the virtual representation of a real world physical feature (object or system). Even though 3D point cloud can be more suitable to depict the real world, it requires HPC. 3D model representations help applications to quickly handle and navigate data comparing to point cloud. However, the 3D model construction of every physical feature usually necessitates expensive time and labor resources to organize and extract the features outlines by interactive manual operations. Recently, deep learning is used to derive semantic classes necessary for 3D modeling by automated classification and segmentation.
PinSout (Point-in Space-out) is a new framework to automatically generate CityGML LOD4 from raw 3D point cloud data by using PointNet. This framework extracts each object required for 3D indoor space modeling from point cloud after learning the deep model with the annotated dataset of Standford Building Parser. After the semantic segmentation, it computes the contour of an object using PCL to augment each spatial indoor model. Finally, the extracted objects are stored into 3D CityDB and provided as CityGML LOD4 data.