In my area, it also appears that the detection rate is fairly good. In my brief look through I found one building where there is no building and a few buildings that the AI did not find. The rotation issue was fairly common as well as a general offset from the imagery (This may be on the Mapbox side, I don't know). The footprints get the general shape of the building mostly there, but struggle with smaller sections. Overall, the tool seems to be good at saying "There is a building here and it's about this big" but is less successful at identifying the details of the buildings.
On Fri, Aug 2, 2019 at 8:17 PM Jmapb <jm...@gmx.com> wrote: > > On 8/2/2019 7:24 AM, Darafei "Komяpa" Praliaskouski wrote: > > Here's a demo by azavea showing how 125 Million AI-mapped buildings > > relate to 33 Million buildings currently in OpenStreetMap in the same > > region. > > > > https://demos.azavea.com/building-footprint-comparison/#4.4/38.67/-93.93 > > Thanks, this is interesting. > > Browsing the areas I'm most familiar with in NYC (where we had a > building footprint import from city data in 2013), I find that the > buildings that are detected by Microsoft but missing from OSM are mostly > either already demolished buildings (currently mapped as > landuse=construction), small outbuildings, or simply nonexistent. In the > cemetery, Microsoft has picked out a few of the more impressive tombs > and mausoleums. And in one case Azavea seems to have failed to notice a > building that's been mapped in OSM for years (since the import.) > > Browsing more rural areas in upstate NY, I see thousands of unmapped > buildings, mostly houses, that Microsoft successfully detected -- in > some cases despite tree cover. The sizes are pretty good and the > footprints are okayish. Oddly some of them seem slightly rotated, maybe > a trick of the shadows. > > J > > > _______________________________________________ > talk mailing list > talk@openstreetmap.org > https://lists.openstreetmap.org/listinfo/talk _______________________________________________ talk mailing list talk@openstreetmap.org https://lists.openstreetmap.org/listinfo/talk