[Osmf-talk] google Open Buildings usage request

b.kilhu+ytrwslnopdfujmrtjzkvsxweizfjncifmzdwwypiihjzdikulpnvql+- at gmail.com b.kilhu+ytrwslnopdfujmrtjzkvsxweizfjncifmzdwwypiihjzdikulpnvql+- at gmail.com
Fri Jul 30 22:51:41 UTC 2021


See The Machine Learning Enabler and MapWith AI (RapiD) road
prediction from Facebook:
https://www.hotosm.org/updates/the-machine-learning-enabler/
https://mapwith.ai/


On Sat, Jul 31, 2021 at 12:05 AM Craig Allan <allan at iafrica.com> wrote:
>
> Dear List,
> As an African mapper, working mainly in Chad, DRC and Kenya, I am
> perpetually scared that some well-meaning AI professional in a far-off
> wealthy country is going to proudly overwrite Africa with dirty data.
>
> In that regard I note that this latest 'Open Buildings' building
> recognition project seems to be good work, at the current level of AI
> image recognition. That is very cool and the field of work is
> fascinating.  But the quality is not yet good enough for OSM purposes.
> The authors of the paper are open about this and are working hard to
> improve their accuracy.
>
> One day, when AI image recognition consistently produces work that
> cannot be distinguished from the work of an experienced human then we
> should certainly look at using AI to classify features.  Until then I
> vote *NO!!!!* to any overwrite.
>
> An African example of the worst case already happening is Cameroon,
> where an AI enthusiast blanketed the country with bad vegetation
> coverage data that nobody will ever bother to fix, because it looks
> 'done', whereas it is not. It is dirty data, which far worse than
> blank.   If the same thing had happened in Germany there would have been
> a riot on this channel.
>
> I also find that if you politely write to mass editors to request
> correction, there is no recourse. The mass editor can't fix the errors
> because that involves hand editing millions of points and ways and they
> don't have the resources to do that. The alternatives are to dump the
> whole import, which is impossible, or fix it yourself.
>
> I would be quite happy with AI produced work being loaded up as a
> digitising aid on another layer, like OpenTopo, and I expect that that
> such limited use as reference-only material would probably be easier to
> licence.
>
> Craig Allan
> osm: cRaIgalLAn
>
>
> On 2021/07/30 09:31, Enock Seth Nyamador wrote:
> > ...
> > In Ghana alone, there is enough of very bad mappings from remote
> > mapping in recent years that we wouldn't want to see more added.
>
>
> _______________________________________________
> osmf-talk mailing list
> osmf-talk at openstreetmap.org
> https://lists.openstreetmap.org/listinfo/osmf-talk



More information about the osmf-talk mailing list