[armchairhacker]

How feasible would it be to generate high-quality guassian splats of everywhere, starting with big cities, like Google Maps street view and 3D view?

Then you can make video games and interactive experiences in real-world locations. I doubt the collision handling is there, but you can at least start with something like Microsoft Flight Simulator with low-flying drones.

It would also be great for training AI to generate realistic-looking scenarios. Besides playing and working (ex: VR) in real places, you can in very-realistic liminal spaces. (And it's training on public areas, so less ethical issues.)

[akanet]

the collision handling part is actually not that hard to back into from the splat data. i think HQ splats of everywhere is something we'll start seeing in the next decade! we haven't quite solved scaling issues, LoD, and streaming for splats but the early work is promising

[jg0r3]

Could you direct me to resources about collisions and splat data?

I have an extreme interest in deriving measurements from splat data of vegetation, as it tends to reconstruct thin planes like leaves far better than other traditional SfM techniques.

[ipsum2]

They already are using NeRFs for Street View. (NeRF is similar to a Gaussian splat with different rendering trade-offs) https://blog.google/products/maps/sustainable-immersive-maps...