All right, good afternoon everybody for those of you. I haven't had the pleasure of meeting yet. I'm Tim Smith I'm a program manager at Microsoft managing both our editorial programs as well as our machine learning and computer vision stuff so buildings buildings buildings, but Canada this time so a quick

Click a little quick little refresher if you're not familiar with our buildings projects And I know a lot of people have loved to name-drop that Basically took semantic segmentation approach this you know, it's a Looking at every pixel in the image in this case aerial imagery and ultimately it's saying, you know Is this a building or is it not a building?

And we did that for the US and that was great and it looked kind of like this and There's a whole bunch of post-processing goes on top of that to make sure that we have good quality It's a it's kind of an interesting problem because that's mantic Segmentation will kind of give us nice Blobs, not the kind of stuff that we want on a map. So we

did develop this post-processing pipeline that will Look at all those edges Create a contour function and ultimately that's what makes nice square buildings at the end I think there's still an outstanding bug about water towers Apparently all the water towers are square now

But how are we able to do all this? You need training data for this stuff. So back in Long ago time before I joined the maps team Microsoft did attempt to to do the whole let's let's make our own. Let's make our own maps So we actually had a whole bunch of labeled labeled building data that was an excellent seed that you know, we made sure it matched up with the imagery and we're able to

Use that as a our training and our baseline that worked great for the US now the the Canada question comes up we actually ran a partnership with Statistics Canada over this because Just simply porting a model that works somewhere there's no guarantee that it's going to behave You know appropriately in a new location

We did attempt that and of course the results were definitely subpar This is due to changes in imagery quality changes in geography different patterns So what the with Statistics Canada was actually running at the time and I think that they've since released this That we kind of got an early share of this was the open buildings database or ODB 2.0

Which they basically went through, you know Statistics Canada is the federal statistics Organization they went through all the different provinces and major cities Hooked up with all their GIS departments and said give us all your buildings and they put it into a very large zip file

That was about 400,000 buildings or so if I recall correctly, that's what's in my deck. So I'm going to assume that's correct Which is an excellent point for us to start because it like without that we're basically at square zero, right? We're gonna have to if the training data doesn't apply. We have to get new training data

and That's you know, a difficult challenge. Otherwise, so partnerships like this are excellent Then there was one additional follow-up that came out of that is there still, you know market specific problems It's one thing to have the known goods in your training data sets But we also had to deal with you know what is not a building that is unique to Canada and we had a bunch of this for the US but

Canada apparently has different types of fields with slightly different degrees of brown And different widths of the lines for which we sow the fields So, you know feel the the images on the right. I don't have any of our bug images sadly I'm sure that would have been would have been funny But suffice to say they kind of look like Sonic the Hedgehog or something

So I mainly want to kind of give some context of what it's like to translate one of these Problems in these projects from one market to the other in the end. This actually did work quite well We were able to release. I think we identified about 12 million buildings across

Across Canada the precision ended up being quite high after a few iterations kind of, you know, taking out those those hedgehog problems And this is also hosted much like our our Canadian building so This stuff looks lovely. I hope that comes out well enough on the slides This has been made available. I think if you saw the rapid demonstration

This is also included in the building contributions that we have for that. So if you guys are enthusiastic about Quickly applying ML resources into OSM. Make sure to check that out. I think this is Downtown Montreal actually and you can actually see the the before and after these are Initially some very old OSM buildings that were kind of poorly mapped and we went in there and we actually saw that

you know in these these scenarios were actually able to extract both better aligned and More articulated buildings, so I think my five minutes is just about up so I will pass it on to the next folks I hope that is a you found that interesting. Thank you very much for your time