MErenB/img-to-road-graph

U-Net road-network extraction from SpaceNet-3 Khartoum satellite imagery: mask -> skeleton -> routable GeoJSON graph, with an OSM coverage diff. FastAPI + Vue 3 demo. # Khartoum Road Graph Extractor Extracts a routable road network from satellite imagery: a U-Net segments the roads, the mask is skeletonised and traced into a graph, and the result comes back as GeoJSON in WGS-84 — then gets compared against what OpenStreetMap already has for the same tile. ## What it does **Model** — U-Net with an ImageNet-pretrained ResNet-34 encoder, trained on SpaceNet-3 AOI-5 Khartoum PS-RGB imagery with road-centreline labels buffered into masks. **Inference** — sliding-window over the tile (512 px windows, 128 px overlap) → hysteresis threshold + morphological cleanup → `skeletonize` → junction/endpoint tracing → `networkx` graph → GeoJSON. **OSM diff** — an OSM reference layer (Overpass, clipped to the tile) is overlaid on the extracted graph, with a diff layer showing what the model found that OSM does not have. Several demo tiles sit in areas OSM barely covers, which is where extracting the network from imagery actually earns its keep. **App** — Vue 3 + Tailwind single-page frontend (Vite) over a FastAPI backend. The demo tiles all come from the spatially-blocked test split — the model never saw them in training — and their results are precomputed so the page loads instantly. Tiles can also be stitched into multi-tile clusters. ## Layout train.py training loop models.py U-Net / encoder setup dataset.py SpaceNet tiles -> tensors create_labels.py centreline GeoJSON -> raster masks app/inference.py sliding-window prediction app/graph_refine.py mask -> skeleton -> networkx graph -> GeoJSON app/osm_ref.py Overpass fetch, clipped to tile app/osm_diff.py extracted-vs-OSM comparison app/mosaic.py multi-tile stitching app/precompute.py caches demo tile results app/main.py FastAPI app app/frontend/ Vue 3 + Vite SPA ## Run locally Backend: ```bash pip install -r app/requirements.txt # plus torch/torchvision …