Geo-typical Synthetic Labels for Building Detection

This dataset supports research in building segmentation within remote sensing using deep learning. It targets the challenge of generalizing segmentation models across diverse geographic regions, where variations in building layouts, types, and sizes pose significant domain gaps. The dataset integrates five widely-used real-world datasets (e.g., INRIA, DSTL) with two synthetic datasets created specifically to replicate geographic regions such as Columbus (Cbus) and Chicago (CHI). Using procedural modeling and physics-based rendering, the synthetic datasets incorporate geospatial data (e.g., street networks) and domain randomization to emulate urban patterns seen in real-world environments.