[OSM-talk] AI detecting of buildings Idle thoughts

[Andy Townsend]

On 16/08/2018 12:35, Rory McCann wrote:
> What's funny is that this import was (according to the changeset
> comment) based on "DigitalGlobe extracted building data". A straight up
> import of the original building geometries would probably be (i) less
> contentious (since a building is a building is a building)

I think the problem here is that "a building is a building is a 
building" isn't really true.  There are many different kinds of 
structures in many different parts of the world, and many different 
light conditions, and different ground surfaces.

We saw this problem with the Facebook "roads import" in Egypt - whatever 
software they were using was detected many sharp edges in imagery 
(walls, canals, etc.) as roads.  If you train whatever you're using to 
detect stuff in one environment and try and use it in another 
environment it's going to get things wrong in unpredictable ways, and 
that's not really going to be obvious if you then (as here) estimate 
residential areas based on extracted buildings.  I can, however, see the 
sense of trying to do that (estimate residential areas based on 
extracted buildings) - in some areas actual buildings are likely to be 
ephemeral, but usage of areas not so much.

Best Regards,

Andy