3D Building Segmentation Using Deep Learning
Computer-vision based technology is becoming more prevalent within the AEC industry, for example, as applied to autonomous drone navigation or AI-assisted architectural design software. In order to function optimally when performing tasks like site navigation or 3D scene understanding, AI technology must recognize higher level architectural features.
As a response, we present a new deep learning based pipeline that rapidly converts previously un-labelled 3D building geometry into richly labelled assemblages of architectural features. Converted models are not only better representations of reality but also provide superior and more detailed building information better suited for computer vision related tasks.
Technology
By combining both parametric modelling tools, dataset preprocessing scripts, and existing Deep Neural Networks models, our pipeline makes it easy to segment thousands of 3D building meshes in a fraction of the time it traditionally took to manually complete this task via traditional software methods.
3D Part Segmentation via Deep Learning
• Deep Learning models were taught to autonomously identify and segment a building into its main architectural features.• This method can accurately segment thousands of buildings at a rapid pace.• Learned features include roof, windows, rooftop units / chimneys, walls, street-facing facade, foundation, additions.• The examples below illustrate how our model segmented point cloud building representations into their primary architectural components.• top row = row-style houses, bottom row = mansard-style houses.
3D Typology Classification
• Beyond segmentation, Deep Learning models were also taught to classify the typological style of buildings• For example, our model can differentiate between row-style houses and mansard-style houses as shown below• Our model correctly classified the top building as being a Mansard-style house with 96.4% probability• Our model correctly classified the bottom building as being a Row-style house with 99.9% probabbility
