Especially true when you realise there is no 13 in that street.
7 Likes
You can do this indirectly by looking for islands of missing UPRNs in an otherwise well-covered area. I do have a heatmap layer that does this, but might need some work before making public.
That sounds about right. In the examples above, the AI seems to be recognising roof outlines rather than building footprints, so the results are poor for āangledā images (canāt recall the correct term), even assuming good alignment of the imagery.
The Ordnance Survey have mixed up over 13 on occasion. Unusually for Richmond there is no 13 Somerton Avenue and the higher numbered houses are out by 2.
Iāve just edited my original post to add in a poll - Iād like to see peopleās opinions on whether youād find it acceptable for poor quality AI buildings such as these (in need of splitting) to be deleted without immediate replacement, or if you think they should gradually be deleted as they are split or replaced. (We seem to all agree that they are bad, but whether we should perform mass deletions or gradual improvement/replacement hasnāt really been discussed yet.)
Personally my vote is to prioritise deletion, so that we can start afresh and donāt give the illusion of completeness that has been alluded to.
1 Like
I donāt see the problem with using address nodes as a quick way to improve building=terrace blocks, often buildings in dense cities have multiple addresses with different doors and overlapping floor plans or a single property lot designation in the cadastre data and so address nodes is the way to map them. An address node also provides more utility to the end user as itāll be close to the entrance door rather than in the middle of the building. In British suburbia where a building is tiny, the front door is generally obvious but even there there are exceptions where the front door is on the side of the building.
1 Like
I would point out that redrawing the buildings from imagery would have the advantage that the result is not encumbered by Microsoft intellectual property.
3 Likes
I donāt bother deleting them en-masse because they are technically correct, just low resolution. I do delete them when working in an area instead of attempting to split them or whatever, just because itās easier.
1 Like
I added a bunch of such buildings in Elmira NY. Iāve corrected a bunch of them, but there still remain some to be fixed. Itās on my long-term todo list.
Is there any encumbrance? Before raising doubts, shouldnāt we cite the license?
100% donāt delete someone elseās additions without consulting them. Itās easy to delete and hard to add. Having your edits deleted is very discouraging. Our community is small enough that we donāt want to discourage anyone.
3 Likes
This is as old news as the MS building dataset is, and has been discussed many many times before. The data is licensed on ODbL terms which while nominally compatible with OSM, creates the two issues:
- we canāt change the OSM distribution licence without permission from MS (or deleting all their data), which would be surprising for most users that have agreed to the contributor terms that assume that there is a democratic process that will govern any such change.
- MS could require attribution any time OSM is publicly used (actually we should really be doing that without them asking).
A further point I would note that I consider it in general bad policy to allow non-sublicensable third party data to be imported, but that ship sailed a long long time ago.
Sure, but if the additions are āAIā data that the mapper hasnāt bothered to square or align, was it really hard to add? Thatās basically shovelling data into OSM.
8 Likes
Sure without but. If a mapper is importing or creating poor quality data, then rather than delete it, you should talk to them. If you just delete it, 1) theyāll get angry and 2) they wonāt learn to be a better mapper.
3 Likes
I doubt the AI is discouraged from editing either. Itās far too dumb for such feelings.
2 Likes
Have you encountered these mappers?
Have you talked to them?
Have they responded?
I havenāt encountered any of these mappers, sorry.
When Iām doing this sort of work I tend to use JOSMās conflate tool to try to preserve a bit of history after completely redrawing the area with Building Tools etc.
Although this has mostly been when the initial poor traces are by enthusiastic new users rather than machine learning.
2 Likes
Missing poll option for me:
āItās ok to delete poor quality AI buildings without replacing them immediatelyā but āthe examples given are NOT a poor enough quality to justify thisā
The sort of thing that might be a poor enough quality for me would be:
- things that are obviously not buildings (umbrellas, temporarily parked vehicles, etc.)
- ābuildingsā that less than 50% match any actual buildings present (much less common now than 12 or so years ago). While there are things that need tidying up with your examples, they do mostly match āwhere there is a buildingā.
10 Likes
I donāt think we even need a consensus check for that, deleting building ways that do not correspond to actual buildings was always correct.
The poll is asking about specifically terraced and semi detached houses.
Although, I am experienced enough with StreetComplete at this point to know to undo the building address and add two+ address nodes if I come across this kind of situation.
4 Likes