🏙️ Pixels, Points & Polygons — Building Vectorizer

Predict building outline polygons from a single aerial RGB tile using the Pix2Poly image model from The P³ Dataset (Sulzer et al., 2025). A DINO ViT-S/8 encoder feeds a transformer polygon decoder that autoregressively emits vertices; an optimal-transport head connects them into closed polygons.

Inputs are resized to 224×224 (the model was trained on 25 cm GSD nadir tiles). Weights: rsi/PixelsPointsPolygons · Code: github