RoadGIE: Towards A Global-Scale Aerial Benchmark for Generalizable Interactive Road Extraction

Researchers introduced WorldRoadSeg-360K, a global aerial road dataset with 366,947 high-resolution images from 38 countries and 223 cities.

They also proposed RoadGIE, an interactive road extraction framework that uses clicks and scribbles to better preserve road connectivity.

The method improved segmentation accuracy and topological consistency, even with a lightweight model.

The work broadens road mapping research with a larger, more diverse benchmark and a more structure-aware interaction design.