oddly enough a relevant arXiv paper was released today: https://arxiv.org/abs/1605.08323
Also how does it compare with https://github.com/mitmul/ssai-cnn
Also how does it compare with https://github.com/mitmul/ssai-cnn
Looking at that repo, the main differences are:
* that code applies to a specific (massachusetts) data set, while DeepOSM uses nationwide USA NAIPs trained against OSM extracts
* they seem to have implemented the full Mnih CNN - DeepOSM just uses a single layer as is
* Mnih didn't touch on using the infrared band
In general, that project seems more mature in replicating Mnih than DeepOSM. Our next goal is a website to show errors in OSM data to be corrected, rather than finish making the net deeper.
edit: added links to paper and code https://github.com/trailbehind/DeepOSM/commit/8053057635a1a0...
> Our next goal is a website to show errors in OSM data to be corrected
Do you mean feeding the data to one of the existing projects that crowdsource fixes based on lists of potential errors, or a new custom site?
Here's the app work: https://github.com/trailbehind/DeepOSM/compare/feature/deepo...
https://github.com/osmlab/to-fix
or
https://github.com/maproulette/maproulette2/
to help with evaluating/fixing of the errors you find.
For example: http://labs.strava.com/routing-errors/#250/17/-46.42546/-21.... gives a fairly strong indication that something is navigable on bicycle or foot; even if OSM hasn't mapped it yet.
My company Gaia GPS has millions of GPS traces from hiking and similar outdoorsy activities, so I thought to combine that data (though footways may be more challenging than roads to classify). OSM trace uploads and Strava both seem like good data sources to do this work too.
Would you be open to working with them to deliver the potential missing roads; so that you don't have to reimplement a lot of the basics?