TernausNetV2

Why we included this project

This repo grew out of a second-place finish in the CVPR 2018 DeepGlobe Building Extraction Challenge, and the network definition, pretrained weights, and a demo notebook are all here, so you can run inference on your own satellite tiles without rebuilding the architecture from scratch. The model takes the original TernausNet and swaps in a WideResNet-38 encoder with 11 input channels, which lets it combine RGB with multispectral bands, and it outputs both a pixel-level building mask and a boundary mask that a watershed transform uses to separate buildings that touch. That makes it a practical reference for anyone pulling building footprints out of overhead imagery for urban planning, population monitoring, or similar mapping work. Just keep in mind it is a research codebase, not a maintained service, so treat it as a strong starting point rather than something to drop straight into production.