SLOD2+WIN: semantics-aware addition and LoD of 3D window details for LoD2 CityGML models with textures

In many urban planning and visualization applications, it is crucial to have 3D window details. However, the process of acquiring and reconstructing them can be challenging. Therefore, in many 3D city models, buildings often lack 3D windows. Instead, their building façades are usually represented using 2D planar textures (i.e., LoD2 CityGML models with textures). To generate a 3D façade from 2D images, current methods often need to carefully design various grammars to achieve desired results, which can be tedious. A useful property of building, the window semantics information, is also lacking. The main contribution of this paper is proposing a semantics-aware method for addition and LoD control of 3D window details to LoD2 CityGML models with textures (namely SLOD2+WIN). A deep learning-based two-level window-pane detection is introduced, and the detection results are then processed and adjusted to generate 3D windows and add to the building models. Unlike other methods, the semantics (i.e., frames and panes) are considered for adding window details. We also propose and integrate a LoD scheme for 3D windows following the same concept as the LoD in CityGML. The tedious efforts of reconstruction or grammar creation can be reduced in our method. Only the information present in the texture itself is extracted, and the shape and pattern information of windows are obtained and adjusted from the detection results in an efficient and unsupervised manner to achieve neat window parsing. Specifically, clustering-based window/pane alignment, neatness-based window image voting, grid-based symmetry, thickness filtering, and fitting-based window-top modeling are proposed. To demonstrate the effectiveness and usefulness of SLOD2+WIN, experiments on several datasets of 3D cities are conducted to add 3D window details with the novel LoD schema, illustrative applications in architectural visualization and urban planning are also showcased. This paper extended its conference version.