This repository contains several implementation demos of deep learning based building extraction with tensorflow.
-
Single-pixel based segmentation: simple CNN & Residual Network.
-
Fully convolutional network, based on the paper Fully Convolutional Networks for Semantic Segmentation.
-
U-Net, based on the paper U-Net: Convolutional Networks for Biomedical Image Segmentation.
A small example data from ISPRS is provided in data/vaihingen-trial-data-raw. A cropping script is provided as data/crop_tile.py, which can crop a big tile to small patches.