Learning Building Extraction from Crowdsourcing Maps

For developing countries and regions, building maps are very scarce but highly desired for various applications. Crowdsourcing maps are often the only map data available, which are notorious for inconsistent quality. We leverage a new approach to learn to map buildings from low quality maps.

We collect
building footprints from OpenStreetMap in two cities, Yaounde, Cameroon and Kano, Nigeria. We use worldview-2 satellite color images at 0.5 m resolution. We compile a training set containing 1060 images of 500*500 pixels and building masks. There are many inconsistencies between maps and labels (see examples below). Note that misalignment errors have already been reduced by a simple procedure.

yao1  yao2  yao3  yao4
kano1  kano2  kano3  kano4

We trained a network with the dataset for around 50 hours on a single GPU. The network is applied to areas where almost no buildings have been mapped. Below is the extraction result on an image of 10,000*10,000 pixels from Yaounde, Cameroon. No pre- or post-processing is performed. Extracted buildings are marked in transparent red. This image takes around 1 minus to process on a GPU.

Two images below correspond to the Kano area, covering the majority of the city and its surrounding areas. Building boundaries are marked in red. The total size is 75,000 * 36,000 pixels. Left shows the available OSM data, which are used in training. Right shows the output from the trained network. 


A few zoom-in views.




If you have questions, please contact yuanj at ornl dot gov.