GAST: Geometry-Aware Structure Transformer

Maxim Khomiakov, Michael Riis Andersen, Jes Frellsen; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2024, pp. 785-793

We present GAST, a novel model for realistic building delineation trained using noisy data designed for real-life applications. While most popular methods today rely on some form of semantic segmentation, the core task of interest is not the points of the interior of the building, but rather the sequence of points surrounding the outer hull - The most sparse set of points encapulating the geometry of the building. Our method works end-to-end, removing the need for post-processing whilst demonstrating generalization across large geographical differences. We compare our method to state-of-the-art complementary works and demonstrate that our model outperforms the baselines in a variety of circumstances, and in all metrics relating to polygon fidelity. We release the dataset and model checkpoints at https://huggingface.co/datasets/anon345/ERBD