hugging-apps/pixels-points-polygons-building-vectorizer

Pixels, Points & Polygons — Building Vectorizer

Interactive demo of the Pix2Poly image model from *The P³ Dataset: Pixels, Points and Polygons for Multimodal Building Vectorization* (Sulzer, Duan, Girard & Lafarge, 2025).

Upload a nadir aerial RGB tile and the model predicts closed building outline polygons, drawn as a vector overlay. The pipeline is a DINO ViT-S/8 encoder → transformer polygon decoder (autoregressive vertex generation) → optimal-transport permutation head that connects vertices into polygons.

• —Weights: `rsi/PixelsPointsPolygons` (checkpoint v4_image_vit_bs4x16)
• —Code: github.com/raphaelsulzer/pixelspointspolygons

Inputs are resized to 224×224; the model was trained on 25 cm ground-sampling-distance imagery, so real-world aerial tiles at similar scale work best. Only the image modality is served here (the LiDAR / fusion variants need Open3D-ML and custom CUDA ops).