Press "Enter" to skip to content

This is how Facebook’s AI maps the planet

Press Release

Facebook has developed a project called Map With AI that is able to identify roads from raw satellite images without any assistance from humans. If this is accurate, maps could be generated from satellite images automatically.

Many people are not yet aware that Facebook, while being a huge social network, has its hands in a lot of pies, and the company is involved in many different research projects that are not necessarily related to social networking. One of these is Map With AI, an AI project that aims to complete maps without having to put in manual work.

Today, maps are created manually and are accurate for developed countries. However, there are many parts of the world where roads have not been mapped properly, and it would take an enormous amount of human effort if maps were to be created for them. There is therefore major value in this project if it could accurately turn a satellite image into a useful map.

Data and process

The website OpenStreetMap was used to download the data used to train the AI, along with satellite images. Then, unsupervised learning was used, where the AI was left on its own to try to identify roads in the image. To train the AI more effectively the researchers only kept images that had lots of roads within them. Then, once a decent amount of images had been fed to the AI, it was used to try to identify roads on new images.

Already the project has had some success, as it was used to map roads in Thailand. If humans had done this task, researchers estimate it would have taken a team of 100 people over three years.

How you can help

The AI is still very far from perfect: it can easily identify rivers or beaches as roads. To really help train the AI, it’s best if a human can oversee it to correct obvious errors. You can also look at routes that have been generated by the AI and then connect them to roads that have been identified in OpenStreetMap.

Anabella Greene