Norwegian Microsoft ML Buildings - More harm than good?

[Johannes-Andersen]

Hey!

Has there been any analysis of false positives, and are there any benefits to using GlobalMLBuildingFootprints in Norway? Given that Norway is a coastal country, it seems like false positives around docks are even more prominent.

Also, based on a few random samples, the data appears to be from around 2015, which means many footprints refer to buildings that have since been demolished. Is newer data available? If not, I would have expected OSM editors to have mapped most buildings by that time making the data even less reliable.

I’d be interested to hear whether it’s worth keeping this dataset enabled. Flying around my local area in Rapid Editor and reviewing the ML building recommendations seems to yield very few valid suggestions. That said, this could be confirmation bias, and it might be more valuable in other parts of the country.

My main issue with the dataset is that there doesn’t seem to be a way to remove features or provide feedback so that bad, non-existent buildings get filtered out over time.

Boat as buildings - example 1.

Boats as buildings - example 2.

Out of date - example 3.

Thanks,
Johannes

[Johannes-Andersen]

It seems that there is a newer dataset available than 2015 as I mentioned.

A lot of the buildings I find that have since been removed are timestamped as 2015 in explore.overturemaps.org.

So my question is: is there logic to remove buildings when a newer dataset no longer contains a building that an older dataset introduced? That would allow newer (and hopefully better) ML models to correct older false-positive mistakes, however at the cost of potentially removing true buildings if the newer model produces false negatives.

I’m not sure how to verify whether the 2015 buildings I’m seeing in Rapid and Overture are still present in newer releases or have already been removed, so I’m not sure whether this approach would help in practice.