Annotated point clouds of buildings: a segmented dataset of single-family houses
The dataset contains 2.904 geometries of single-family houses in the form of annotated Point Clouds, and was developed in order to train 3D Generative Adversarial Networks with architecturally relevant data. More specificaly the geometries are segmented within 3 classes: wall, roof, floor. The points of the point clouds are saved in .pts files while their labels are saved in .seg files.
Steps to reproduce
The creation of the dataset was done in a semi-automated way and in two stages: a) Creation of module geometries that represent building components in Rhinoceros3D, and b) Conversion of the geometries into point clouds using the Cockroach plug-in for Grasshopper. Specifically, 25 wall modules were designed for the generation of building series. Then, 35 roof modules were created and combined with each wall module respectively. Data augmentation methods were applied to maximize the size of the dataset: the modules were scaled in 3 ranges, and rotated 90 degrees for a wider feature space.