Wrong city is detected

[golibali]

The bug

I live in the countryside of Budapest, in a small town. All the photos what I take in the town is registered by Immich to the town next to us ( there is a hill between us) I checked some of the images coordinates and they are in the town (eg: 47,5767277777778 N, 18,8796 E) and is registered as it is in the Pilisszetivan town.

The OS that Immich Server is running on

Ubuntu 22.04

Version of Immich Server

v1.115.0

Version of Immich Mobile App

1.115.0 build.175

Platform with the issue

• Server
• Web
• Mobile

Your docker-compose.yml content

Same as recommended here: https://immich.app/docs/install/docker-compose/

Your .env content

Same as recommended here: https://immich.app/docs/install/docker-compose/

Reproduction steps

1. Take an image in town
2. Mobile app uploads
3. Immich server tag with wrong town label
   ...

Relevant log output

No response

Additional information

I found this ticket: #4745
I this ticket the https://www.geonames.org/ was mentioned as Immich gets info from there. I checked the wrong "Plilisszentiván" town and the rectengle not overlaps with the geo locations of my images. The right town name would be "Nagykovácsi". There are 3 of them on this page, but they are almost covering the same place. Could this be the root issue, that there are multiple entities for one town in the geonames database?

[SanoKriss]

I have same problem with another location 48,1945059722222 17,053632 detected as Devínska Nová Ves and should be Lamač

[mmomjian]

The way we handle geo boundaries will be changed in a future release. Currently this is an expected behavior. Thank you

[SanoKriss]

So, tell me please, which method is used? Thanks

[golibali]

Also why do you close this issue as the bug is valid, I mean it tags images with wrong location.

[kvalev]

Immich uses the data from the cities500.txt file. So your picture will be geo located in the closest village with a population of more than 500 (within some radius). You can check the geonames data and correct it if needed.

[golibali]

@kvalev Thanks, that is the issue here. The "Nagykovácsi" town geo location in the cities500.txt is way too wrong, it's more than 12 km far from the real location of the town.
I checked in this, just for the record: https://download.geonames.org/export/dump/cities500.zip

[zackpollard]

This is where it appears on geonames, which does appear to be wrong. You can correct this by going to geonames.org and changing the location of the populated place marker 🙂

[SanoKriss]

Immich uses the data from the cities500.txt file. So your picture will be geo located in the closest village with a population of more than 500 (within some radius). You can check the geonames data and correct it if needed.

Lamač is part of Bratislava populated with more than 7000 people, but is not in that file. I will check it in geonames, they have several sources of data, for example in allcountry.zip Lamač is there. This is problem of "dirty" data and I understand using source file (cities500.txt) but this not solved problem (not only my). Photos geolocation is wrong marked to nearest city/village.

[kvalev]

You can add the population for Lamac here and after the next immich release if you refresh your metadata, the location should be corrected.
The immich solution is not ideal, but it is what it is and it probably is gonna take a while until another one is in place, so you have to work with what you got.

[mertalev]

We're planning on making changes that will make reverse geocoding more accurate. I believe this will also involve using OSM data instead of GeoNames.

[gthrepwood]

I asked someone to extract the exact boundaries from OSM, and without further optimization, db was ~6GB.

I live in the one side of the capital city, and the place if closer to a small village than the center of the capital.

This is a small workaround: set "Budapest XIX" to the photos, what was 10m radius of my location. Yes, it will overwritten, when immich rescan the exif of old photos.

UPDATE asset_exif ae
SET city = 'Budapest XIX'
FROM asset a
WHERE ae."assetId" = a.id
  AND (
    1000 * 6371 * acos(
        cos(radians(47.42560464469522)) * cos(radians(ae.latitude)) * cos(radians(ae.longitude) - radians(19.201789505737757)) + 
        sin(radians(47.42560464469522)) * sin(radians(ae.latitude))
    )
  ) <= 10;

[kvalev]

There already is a proposal for a fix by @marcules without hacky solutions - #9051 (comment) The only thing missing is a blessing from the maintainers to go with the fix.

[gthrepwood]

As I see, the request pending from Apr 24, 2024 and that is an awful lot of time. It seems to be 'better' that the original cities500 solution, but still questionable, if that would be a trusted improvement.

[marcules]

Hey hi, @gthrepwood my proposal is actually only about 2 weeks old, but I made the mistake of commenting it on an old PR instead of this discussion, as, from my point of view, the linked PR actually introduced a regression for anyone taking pictures in the outskirts of cities.

So I'm not sure if you've read the actual proposal further down the comments, because I don't think it's questionable to use the hierarchy of the nearest populated area to determine the "correct" nearest city.

Of course it would be even better to use actual geofencing or even allow the user to designate their own locations as to create "synthetic"/"manual" populated area markers, that would further improve pictures in that region for the user in question; but the proposal is deliberately small in scope as to not impact any plans the maintainers have with real geofencing, so not to impose on their ideas of a "real geofencing implementation" this is as small as possible, but with the desired effect of improving the location data on outskirts.

[tomzorz]

there was also a thread somewhere about a better solution being implemented / halfway done already; or maybe it was in discord? not sure. but it wasn't #9051

[MichelZ]

I've built exactly this now, it uses overture maps and uses a place's geometry/bounding box, not just a point on the map to calculate the place:
Docs: https://immich-reversegeo.github.io/
GitHub: https://github.com/immich-reversegeo/immich-reversegeo

[MichelZ]

I've built exactly this now, it uses overture maps and uses a place's geometry/bounding box, not just a point on the map to calculate the place:
Docs: https://immich-reversegeo.github.io/
GitHub: https://github.com/immich-reversegeo/immich-reversegeo

[gthrepwood]

I tried @MichelZ 's solution and it's more than enogh as I see. on demand database download, easy to install, works as expected.

I checked OSM's database, how to extract city borders, and it's VERY tricly as China, Greece, Hungary have all use different levels for administrative regions.... I would gladly accept, if someone else done that work, and harmonise the databases.

[ClementTsang]

Ooh, I might try this - I have some photos that are smack dab in the middle of a huge city that it currently decides is a city a few miles over for whatever reason.