Urbex Finder

Why

I like abandoned buildings. Finding them takes hours of going through aerial views. I wanted a model to do it for me and hand back a ranked list I would only have to sort through. A first attempt on my own in February 2024, taken up again seriously in 2026 with Claude Code.

How

OpenStreetMap building footprints serve as candidates: no need to sweep the territory, rural France is over 90 % fields. For each one, a window is taken from the IGN orthophoto, at about 20 cm per pixel. The model is a fine-tuned DINOv2, trained on buildings tagged abandoned, disused or ruins.

The overall score is carried by the collapsed ruins, which are easy to recognise. What interests me is abandoned buildings still standing. On those, the first version (a frozen DINOv2 and a logistic regression) topped out at an average precision (AP) of 0.52. The current model reaches 0.73 in cross-validation, and 0.70 on six départements it never saw. The output: a ranking, a map and an interface for reviewing the candidates by hand.