Efficient Procedural Modelling of Building Façades Based on Windows from Sketches

Realistic and delicate window models are in great demand for detailed LOD-3 urban reconstruction. However, current research tends to oversimplify the window models due to the difficulties in describing the complex window types and structure parameters. This work proposes a sketch-based window reconstruction method for procedural façade modelling. With a rough sketch of the target windows, the method can quickly identify the window type and estimate the grammar parameters, eventually producing detailed 3D models with textures. The whole task is adopted within two neural networks: the classification network to judge the grammar category and the regression network to estimate the parameter value. The non-photorealistic rendering technology adopted in the present work can generate abundant training datasets, and meanwhile, the attention mechanism is used to identify the grammar parameters precisely. Experimental results on various data verify the effectiveness and accuracy of the proposed approach.