U-Net ensemble for building segmentation
Course project IKT452 — aerial imagery, ensemble of U-Nets, MapAI-style evaluation mindset.
Task
From overhead tiles, predict building footprints. Useful for mapping agencies, insurance, and urban analytics — competition-style metrics reward crisp boundaries and topology-aware masks.
Technical notes
- Architecture: ensemble of U-Nets to stabilise predictions and capture complementary errors.
- Data: aerial patches with dense masks; standard augmentations and careful train/val splits to avoid optimistic leakage across tiles.
- Alignment: implementation informed by MapAI competition framing (per course specification).
Illustration placeholder
Before/after overlay: RGB tile + predicted mask in colour — the quickest way to make this page memorable.
Takeaway
Classic encoder–decoder vision with a geospatial twist: strong baseline for CV engineer interviews where “show me a segmentation you have actually trained” comes up.