Reproduce “GeoFormer: A Multi-Polygon Segmentation Transformer” on OpenResearch

About this paper

Maxim Khomiakov, Michael Riis Andersen, Jes Frellsen

Researchers at the Technical University of Denmark developed GeoFormer, an auto-regressive transformer that directly generates multiple building polygons from satellite imagery using a single likelihood function. The model achieved an Average Precision of 91.5 and a Boundary Average Precision of 97.1 on the Aicrowd Mapping Challenge dataset, significantly outperforming previous methods.

Key insights

• An end-to-end auto-regressive Transformer, optimized with a single likelihood function, can effectively generate multiple precise vector polygons without complex multi-loss optimization or raster-to-vector post-processing.
• Modeling polygon vertices as spatially dependent tokens allows for the direct generation of sparse, scale-invariant vector representations that inherently capture building geometry.
• The combination of hierarchical image features and sophisticated positional embeddings (RoPE, ALiBi) in an auto-regressive decoder enables robust learning of ordered geometric sequences for multi-object scenes.

Summary by alphaXiv.