Synthesizing 3D Abstractions by Inverting Procedural Buildings with Transformers

Training is fully supervised, We generate abstractions of buildings, reflecting the essential based on a dataset of procedural buildings paired aspects of their geometry and structure, by learning with corresponding point cloud simulations. We develop to invert procedural models. We first build a dataset of various technical components tailored to the generation of abstract procedural building models paired with simulated abstractions. This includes the design of a programmatic point clouds and then learn the inverse mapping through a language to efficiently represent abstractions, its combination transformer. Given a point cloud, the trained transformer with a technique to guarantee transformer outputs consistent then infers the corresponding abstracted building in terms with the structure imposed by this language, and an of a programmatic language description.