No ML. We templated some common architecture patterns and placed things in the world.
I wish I could show you our internal demonstrations of it. Basically we would plot the map in snowdrop using GIS, then raise cuboids from the shapes on the map. Then we would manually try to match google street view with something in our template library, and modify the object to match reality.
It was a very manual process, but we based the map on the real thing so distances should be 1:1.
I've recently taken up photogrammetry mapping with my drone. I'm quite amazed at the textured 3d output you can get from that by capturing buildings just by flying around them. I imagine it won't be too long that we just fly a drone over a city and entirely map it in 3d with very high resolution and detail. It's accurate to within a few feet unless you use RTK and ground stations which can bring you within centimeter accuracy. Lidar is also making things easier/cheaper.
I wish I could show you our internal demonstrations of it. Basically we would plot the map in snowdrop using GIS, then raise cuboids from the shapes on the map. Then we would manually try to match google street view with something in our template library, and modify the object to match reality.
It was a very manual process, but we based the map on the real thing so distances should be 1:1.