OpenStreetMap
Auto over tracing buildings using "mapwithai-dev" plugin for latest "josm-latest.jar"
Posted by ABZ_OSM on 15 July 2020 in English (English).Any level of automation of over tracing buildings, using AI, would be a game changer for OSM allowing us to populate the map with detail much more rapidly.
Over tracing is normally a laborious task. In my case even using a wacom drawing pad to over trace. If you haven’t got a wacom drawing pad, get one, it makes over drawing much much easiear vs using a mouse.
Here is a 13 minute screen cast of my experience of using “mapwithai-dev” plugin with the latest “josm-latest.jar”.
My intention was to discover if I could over tracing buildings in Aberdeen, Scotland, Europe.
The guidance on how to do this is posted at this page https://gitlab.com/gokaart/JOSM_MapWithAI/-/wikis/Testing-the-development-version
If anyone can shed any light on how to make this plugin useful, for specific locations, that would be very helpful,
thank you
Comment from SomeoneElse on 15 July 2020 at 17:31
How do you handle the offsets between “detected features” and reality? When I last looked at RapiD the imagery used seemed to be the one with the worst offset of the ones available to OSM editors. Usually when adding anything to OSM I’ll align based on a mix of imagery, GPS traces (especially where there are lots of them) and other sources - OS OpenData StreetView can be good for this in the UK.
There’s also the problem of detecting “features that are not features” - my last quick test found lots of detected “service roads” in Mansfield that were either not really useful to map, weren’t really service roads (and were all offset from reality). This is a little unfair- in that area all the real roads have been added so anything “new” detected is obviously going to be rubbish; if used somewhere where there are real major roads to detect that wouldn’t be the case (although as various OSM communities have found to their cost human checking of each piece of data as it is added is still very much needed).
Comment from Sanderd17 on 16 July 2020 at 06:11
I have worked a bit with MapWithAI to map roads in Peru.
Manual checking is still very much needed. Even if the alignment doesn’t matter that much (when there are no other features present), MapWithAI seems to struggle with tight bends and foiliage covering the streets (where a human would just connect it through).
That said, the plugin did help a lot with lowering the amount of work. It’s far less work to just scroll over a road, and touch up a few points, than having to draw it completely by hand.
Comment from ABZ_OSM on 16 July 2020 at 13:37
Hi @Sanderd17
Thanks for your reply. That is useful to know.
Which tool did you use, and since I got no where with this, any chance you could do a screen cast to show what you are doing, or describe how you did this.
Comment from Sanderd17 on 16 July 2020 at 15:36
I used the older version of the plugin, not this alpha release. The old version can just be installed from inside the JOSM plugins.
But it only supports roads, hence it’s only useful in unmapped areas.
Comment from Zkir on 17 July 2020 at 11:15
Hi, I’ve not got, have you managed to make the plugin recognize buildings ?
Comment from ABZ_OSM on 17 July 2020 at 12:57
Hi @Zkir
The diary post above contains a link to a screencast youtube video I made of my experience which spans 13 minutes. Here, https://youtu.be/JtyhdFH39cY
No I was not able to figure out how to get it to overtrace buildings; becasue I do not think it does, for anywhere here in Scotland, Europe.
It seems that Facebook is working on AI recognition for roads, and Microsoft (MS) is on AI recognition for buildings.
I think how this AI works, as far as I know, is MS (or other) would provide suggested AI generated over trace “way” files for buildings in the JOSM plugin or RapidID. Within the JOSM plugin these pre-generated files would be loaded and corrections applied as needed.
There seem to be files available, within the plugin, for parts of California, USA, but these files seemed to me, to have been over traced with ways already.
If it works any other way, then perhaps someone will outline how it does actually work.
If you look at the tail end of the video you will see what I mean