[Talk-GB] Map with AI comes to the UK

[SK53]

I'm using QGiS 3.4 on a laptop with 8 Gb, but I'm sure much less will work
fine. Will send data separately.

I'm not sure Facebook have corrected for offsets in imagery so whereas the
topology may be better positioning may be worse. One of the simple quick
wins from using ML on imagery could be better alignment of multiple imagery
layers (provided one can extract the same types of features from them &
ones which are algorithmically easy to align - there's a huge literature on
this, colleagues were working on
<https://www.sciencedirect.com/science/article/pii/0262885689900012>
matching stereo SPOT images in the '80s).

Using OS Open Roads (which looks to be slightly generalised so some slight
curves are straight lines) and the buffer clipping technique (e.g. as I
used for Northern Ireland townlands) is probably the best way to realign
NPE roads. I currently don't have the disk space to do this easily,
otherwise I would have done).

Jerry

On Sun, 5 Jan 2020 at 15:03, Rob Nickerson <rob.j.nickerson at gmail.com>
wrote:

> Thanks Jerry.
>
> I'd also subsequently discovered the data dump but had not yet got around
> to looking at it. What are you using here to view and work with the data?
> Is QGIS and 6GB RAM sufficient?
>
> I would be interested in Warwickshire if you can extract that.
>
> And yes, we probably are not expecting much for the UK given how well we
> already have most roads mapped. It's a shame it only shows missing roads as
> I suspect it has better geometry of some rural roads in poorly mapped areas
> than us - I still find jagged roads with source=npe.
>
> Nevertheless, the AI stuff is an interesting one to keep an eye on. If
> improvements can be made and additional datasets incorporated, it could
> become a significant aid in the future. For example I wonder if it could be
> good at building detections when combined with other data such as LiDAR
> height data. There is also the prospect of using AI to help find solar
> panels.
>
> Best regards,
> *Rob*
>
>
> On Sun, 5 Jan 2020 at 14:07, SK53 <sk53.osm at gmail.com> wrote:
>
>> Perhaps more useful is that one can download the UK data as a geopackage
>> from
>> https://github.com/facebookmicrosites/Open-Mapping-At-Facebook/wiki/Available-Countries.
>> It's 147Mb zipped in a tar which unpacked is around 400Mb.
>>
>> I've had a very quick look and notice quite a few concentrations of
>> features which are obviously tractor lines in farmland. See this area
>> around Colston Bassett
>> https://wiki.openstreetmap.org/wiki/File:Colston_basset_fb_rapid.png.
>> Apparently such false positives can be marked as such in the editor which
>> ought to improve detections next time round. My suspicion is that things
>> which are actually roads are mainly driveways to outlying houses & farms or
>> farm tracks. Using OS OpenRoads is more likely to help find significant
>> missing adopted roads.
>>
>> Jerry
>>
>> I can potentially provide extracts for individual LAs if people want
>> them, send me an email. I personally found it easier to look at the data as
>> a whole rather than scanning around in the editor.
>>
>>
>> On Thu, 2 Jan 2020 at 22:09, Rob Nickerson <rob.j.nickerson at gmail.com>
>> wrote:
>>
>>> Hi all,
>>>
>>> I just spotted that Facebook have pushed an update to their map with AI
>>> project:
>>>
>>> "For our final release of 2019, we have released 84 new countries for
>>> our AI road data with new coverage in the remainder of Europe, Asia, and
>>> Oceania! AI roads are now available nearly globally."
>>> Source:
>>> https://github.com/facebookmicrosites/Open-Mapping-At-Facebook/blob/master/WHATSNEW.md
>>>
>>> For those who don't know what MapWithAI is check out https://mapwith.ai
>>>
>>> And to try it out in their RapID version of the iD editor:
>>> https://mapwith.ai/rapid#background=Bing&disable_features=boundaries&map=18.60/53.40625/-2.13801
>>>
>>> Just roads at the moment and not that easy to find a suggestion that is
>>> worth adding (at least near me where roads are well mapped) but this does
>>> demonstrate what is possible. Let us know if you have a good or bad
>>> experience with this.
>>>
>>> Would be great to see this extended to buildings but we may have to wait
>>> for Microsoft for that.
>>>
>>> P.S. A happy new year to all!
>>>
>>> Thank you,
>>> *Rob*
>>> _______________________________________________
>>> Talk-GB mailing list
>>> Talk-GB at openstreetmap.org
>>> https://lists.openstreetmap.org/listinfo/talk-gb
>>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstreetmap.org/pipermail/talk-gb/attachments/20200105/a79de22a/attachment.htm>