Reproduce “PolyWorld: Polygonal Building Extraction with Graph Neural Networks in Satellite Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Stefano Zorzi, Shabab Bazrafkan, Stefan Habenschuss, Friedrich Fraundorfer
PolyWorld extracts building footprints from satellite images as polygons, detecting candidate corners and learning how they connect instead of tracing boundaries from raster masks. A graph neural network scores clockwise and counterclockwise connections and refines corner positions. On CrowdAI test data, it reports C-IoU 88.2 versus 73.7 for FFL with ACM polygonization; evidence is limited to this dataset.
Key insights
- Representing building polygons as a set of vertices linked by a permutation matrix, which the network learns to predict directly, enables true end-to-end vectorization.
- Integrating a Graph Neural Network allows for global contextual understanding between all potential building corners, making vertex matching more robust.
- Combining a specific angle loss with a differentiable segmentation loss during training explicitly optimizes for geometrically regular polygons with sharp corners while maintaining overall footprint accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.