Facebook pairs its Map With AI service OpenStreetMap project
Facebook said it will open its Map With AI service to the OpenStreetMap project, which aims to be an open source mapping data source.
In conjunction with the Map With AI service, Facebook also said it would release RapiD, an AI version of the OpenStreetMap editing tool iD.
Mapping data is increasingly critical and Google currently dominates even as Apple, Microsoft and other vendors aim to secure mapping data for everything from search services to directions to local business directories.
Facebook's effort, aside from also benefiting from mapping data, is designed to work with the open source community to map millions of miles that haven't been mapped. Volunteers have been busy cataloging missing roads, buildings and brides on OpenStreetMap or OSM. Facebook said AI has allowed it to map Thailand's missing roads in 18 months--roughly 300,000 miles--and cover more than 90% of missing data in Indonesia. Facebook said traditional means would have taken another three to five years to map.
The Map With AI service includes access to AI-generated road mappings in Afghanistan, Bangladesh, Indonesia, Mexico, Nigeria, Tanzania, and Uganda. The service will cover more countries over time. RapiD is also designed to add and edit road data for volunteers.
While the data in remote locations initially helps disaster recovery and organization like the Red Cross, Facebook could also benefit as these markets develop. Facebook can use better mapping data for Facebook Marketplace and Facebook Local.
How does it work?
Map With AI uses computer vision so machines learn to spot patterns in images via satellite imaging. The AI system is trained to identify possible roads and highlight them. From there, volunteers review and confirm suggestions.
The computer vision system utilizes a deep neural network to highlight probability of a road with colors to indicate confidence.
Key points:
- The AI mapping approach can be deployed at scale and find roads obscured by trees.
- The Facebook AI team has a 34-layer deep neural network model.
- False positives for river beds and other obstacles are checked by volunteers.
- The Thailand mapping project helped Facebook researchers adjust the model and train it.
- Facebook said it collected its training data as a set of 2,048-by-2,048-pixel tiles, with a resolution of about 24 inches per pixel. Tiles with fewer than 25 roads mapped were discarded.
Primers: