GeoFormer: A Multi-Polygon Segmentation Transformer

TL;DR

GeoFormer is a novel transformer-based architecture that learns to generate multi-polygon building segmentations end-to-end, improving accuracy and robustness in remote sensing imagery without complex loss function tuning.

Contribution

It introduces the first auto-regressive transformer model for multi-polygon predictions in remote sensing, simplifying the process and enhancing performance.

Findings

Outperforms existing methods in delineating buildings from satellite images.

Robust against parameter variations and ablations.

Optimizes a single likelihood function for end-to-end learning.

Abstract

In remote sensing there exists a common need for learning scale invariant shapes of objects like buildings. Prior works relies on tweaking multiple loss functions to convert segmentation maps into the final scale invariant representation, necessitating arduous design and optimization. For this purpose we introduce the GeoFormer, a novel architecture which presents a remedy to the said challenges, learning to generate multipolygons end-to-end. By modeling keypoints as spatially dependent tokens in an auto-regressive manner, the GeoFormer outperforms existing works in delineating building objects from satellite imagery. We evaluate the robustness of the GeoFormer against former methods through a variety of parameter ablations and highlight the advantages of optimizing a single likelihood function. Our study presents the first successful application of auto-regressive transformer models for multi-polygon predictions in remote sensing, suggesting a promising methodological alternative for building vectorization.