Home / Data: Planet, Coastlines, Buildings, Admins, Roads / Attribution

New Machine Learning-Detected Roads from Grab Location Signal Data

Nov 17th, 2021

Image-based machine learning models have challenges in detecting roads in highly dense areas due to occlusions from buildings and trees. Meta has been collaborating with Grab, a food delivery company operating in South East Asia. Their machine learning (ML) models trained on the driver location signals provide great value in terms of detecting missing roads in densely populated areas.

As part of this collaboration, we are providing ML-detected roads from Grab location signals as a one-time-addition in our Daylight release. This data is available for free under Creative Commons by Attribution (CC BY 4.0) license and covers 62 cities in Southeast Asia such as Jakarta and Bangkok. We are excited to announce that this one-time release will contribute to OSM adding a total of 240K kms of newly detected roads. Meta map building team is working on making this data available in RapiD.

For this release, ML-detected roads went through the following steps:

• Grab extracted centerlines from ML output masks to create road linestrings.
• Meta conflated Grab linestrings with the roads in Daylight release.
• Meta filtered out linestrings passing through buildings.
• Meta filtered out ferry routes overlapping with water bodies.

While the coverage and quality of conflated output is great, it is worth to note that the data can contain artifacts and weird geometries around some bifurcation points.

📦 Download the ML-Detected roads: grab_ml_roads.osc.bz2 (41.1 MB)