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

TL;DR

PolyWorld is a novel neural network that directly extracts and constructs precise building polygons from satellite images using graph neural networks and optimal transport, improving over existing raster-based methods.

Contribution

Introduces PolyWorld, a method that directly predicts building vertices and their connections to generate accurate polygonal building outlines from satellite imagery.

Findings

Outperforms state-of-the-art in building polygonization

Produces visually pleasing and precise building polygons

Achieves significant quantitative improvements

Abstract

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.