Show HN: Mask R-CNN Neural Network for Mapping Sport Fields in OpenStreetMap

[jasonremillard]

https://github.com/jremillard/images-to-osm

[francisofascii]

This is impressive. I wonder if a similar technique could be used to identify parking lots. The U.S. seems to have an over abundance of parking lots, so it would be cool to back up that perception with real data.

[jasonremillard → francisofascii]

Thanks! It would probably work well for parking lots, but there are a horrifying number of them. I don’t think the OSM community would tolerate such a large import.

[rmc → jasonremillard]

You don't have to do an import, you could just have a MapRoulette challenge. Or some sort of JOSM todo plugin, so people could whiz from one to the next

[toomuchtodo → francisofascii]

https://github.com/trailbehind/DeepOSM

https://news.ycombinator.com/item?id=11808639

[maxerickson → toomuchtodo]

I think the Mask R-CNN used in the linked project is a better fit to parking lots (areas) than DeepOSM, which is aimed at finding road centerlines.

[Mediterraneo10]

I mapped quite a bit of Morocco while traveling through there recently. Interestingly, I find that a lot of sports fields in villages are for both football and basketball, but the basketball structures aren't visible from aerial imagery. Rather you will know they are there only if you actually see the pitch with your own eyes. So, while automatically identifying leisure=pitch might be possible with neural networks, the tagging still needs a human being to visit the site in order to be complete.

[jasonremillard → Mediterraneo10]

This is just for the US North East. Football and basketballs fields are never shared because basketball fields are always paved.

[edshiro]

Very interesting! I recently finished reading the paper on Mask R-CNN and find it's a very neat addition to the R-CNN family.

What are some of the issues that you encountered with the implementation of Mask R-CNN or the training phase for example?

[riordan]

This is awesome and exactly the kind of thing that can give the OSM Armchair Mappers some very accurate superpowers.

It'd be great to see this fed into the OSM Tasking Manager [1] or Maproulette[2] for folks to validate (and then build even stronger training data).

[1]: http://tasks.openstreetmap.us/ [2]: http://maproulette.org/

[maxerickson → riordan]

If I understand correctly the creator has already validated the output for much of the Northeast.

(They've proposed an import: https://lists.openstreetmap.org/pipermail/imports/2017-Decem... )