[crazygringo]

Me too. The problem seems to be mainly that the models are still awfully approximate and the textures are horribly blurry. They're fantastic for a bird's-eye view or flight simulators, but not what you want for driving.

I'm hoping that at some point in the next 10 years, Google/Microsoft come up with a high-res version produced mainly by Street View-type data or similar drone-supplied data. I suspect that basic photogrammetry is the easy part -- that the challenging part is dealing with changing sun position, hourly/daily/monthly weather/seasons, and all of the pedestrians and moving/parked vehicles.

[sbierwagen]

Fun fact: Google street view cars record video-- the widely spaced street view nodes are a small fraction of the imagery they have stored. It should be pretty easy to do higher res photogrammetry with the raw data.

[darkwraithcov]

AI upscaling with diffusion models will probably solve that problem.

[waltbosz]

That's what I was thinking. The satellite photos are detailed enough to see lane lines, and combine that with the street level photos will get you the road configuration.

But I think whatever data GPS software uses is already detailed enough to know lane config, it tells you exactly what lane to be in when making turns.

[tomaaron]

photogrammetry is old. Check out NERF. https://youtu.be/DJ2hcC1orc4

[rzzzt]

LiDAR-mounted drones are available, although a ground-level thingy can also generate neat point clouds in its vicinity: https://enterprise-insights.dji.com/blog/lidar-equipped-uavs