PolyWorld: Polygonal Building Extraction with Graph Neural Networks in Satellite Images
This addresses the need for accurate building polygons in geographic and cartographic applications, representing a novel method for a known bottleneck in instance segmentation.
The paper tackles the problem of extracting precise vector polygons of buildings from satellite images, which is needed for geographic applications, by introducing PolyWorld, a neural network that directly predicts vertices and connects them to form polygons, achieving state-of-the-art performance in building polygonization with notable quantitative results.
While most state-of-the-art instance segmentation methods 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 directly extracts building vertices from an image and connects them correctly to create precise polygons. The model predicts the connection strength between each pair of vertices using a graph neural network and estimates the assignments by solving a differentiable optimal transport problem. Moreover, the vertex positions are optimized by minimizing a combined segmentation and polygonal angle difference loss. PolyWorld significantly outperforms the state of the art in building polygonization and achieves not only notable quantitative results, but also produces visually pleasing building polygons. Code and trained weights are publicly available at https://github.com/zorzi-s/PolyWorldPretrainedNetwork.