S. Zorzi ,  S. Bazrafkan ,  S. Habenschuss ,  F. Fraundorfer (2022)

PolyWorld: Polygonal Building Extraction with Graph Neural Networks in Satellite Images

communication medium

IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR) 2022


While most state-of-the-art instance segmentation meth- ods produce binary segmentation masks, geographic and cartographic applications typically require precise vector polygons of extracted objects instead of rasterized output. This paper introduces PolyWorld, a neural network that di- rectly extracts building vertices from an image and connects them correctly to create precise polygons. The model pre- dicts the connection strength between each pair of vertices using a graph neural network and estimates the assign- ments by solving a differentiable optimal transport problem. Moreover, the vertex positions are optimized by minimiz- ing a combined segmentation and polygonal angle differ- ence loss. PolyWorld significantly outperforms the state of the art in building polygonization and achieves not only no- table quantitative results, but also produces visually pleas- ing building polygons. Code and trained weights are pub- licly available at

research topic

research groups


Computer Vision, Pattern Recognition