This talk was transcribed.
What if a machine could find all the untraced streets and buildings? What if it could trace some of them? This talk will share our experiments using OpenStreetMap data to train computers to recognize buildings and roads in satellite imagery.
Machine learning–specifically, “deep learning”–has made astounding progress. Convolutional neural networks have recently started outperforming humans at labeling everyday photographs. But finding good “training data” is still one of the most limiting factors. The volume of ““ground truthed”” data in OpenStreetMap has enormous potential for training deep neural networks to extract features from satellite imagery.
We’ve been taking convolutional network models that were extremely successful on normal photographs and working on re-training them using aerial images and OpenStreetMap data. In this talk, we’ll share our experiences with this approach. We’ll also walk through our open source code for generating training data from OpenStreetMap and our actual training scripts. Our hope is that others can replicate and build on this work. Finally, we’ll discuss the prospects for using using a neural network trained from OpenStreetMap to (semi-automatically) contribute back to OpenStreetMap.