Positive development for OSM?

Just saw this newsfeed from MIT. https://news.mit.edu/2018/new-way-to-automatically-build-road-maps-with-aerial-images-0417
More of the same, but less ‘popular’ formulated: https://arxiv.org/pdf/1802.03680.pdf

Worth reading too: https://wiki.openstreetmap.org/wiki/AI-Assisted_Road_Tracing

Is this a positive development for OSM?

For what it’s worth, I’m just the messenger…

Depends on whether it’s done right. Previous efforts to apply AI tracing to OSM have being at odds with how OSM works, and apparently suffered from quality issues. But I don’t think it’s impossible to make it community-compatible. Ideally, such technology could simply be another, powerful tool in the hands of local contributors.

As a counterpoint to that, I’d suggest reading some of the threads in the Thai forum, such as https://forum.openstreetmap.org/viewtopic.php?id=57387 . My comment at the start of that thread (“we know what happened last time”) refers to an import by the same people (Facebook) in Egypt, which had to be completely reverted, because anything that looked a bit like a straight line (edges of canals, buildings, walls) got added as a “road”.

Based on what’s been done so far, you’d have to say that the quality of auto- and semi-auto traced stuff is far lower than a human with even just a couple of hours experience of OSM could manage. Even allowing for improvement it’ll never match the level of detail that someone on the ground could manage - you can’t see shop times and opening hours from an aerial photo :slight_smile:

Hey, I am one of the authors of the RoadTracer paper. I am actually very interested in determining how machine learning approaches can be used to improve the OpenStreetMap editing process! I actually just posted here today, regarding a demo consisting of a modified iD editor, before seeing this post.

https://forum.openstreetmap.org/viewtopic.php?id=62138

The idea in the demo is for a semi-automated MapRoulette-style system, where the system suggests tiles containing missing roads detected by the machine learning model. The missing roads are highlighted, but the user still is asked to trace them out. The system eventually should prioritize roads that it thinks are more major/important.

The quality of the missing roads isn’t that great in the demo (but should be good enough for its purpose). I think it could be improved to some extent, especially in terms of reducing noise and getting better alignment, but I think even then full automation might not be desirable because there are still going to be errors.