So an update on this for people who might be interested - I’ve been working on this the last two weeks:
USGS actually had High-resolution Orthoimagery at 1 ft resolution, so I have been training a new network to upscale to 2048x2048 (or 0.5m resolution) using an adversary (one network tries to come up with an upscale, the other network tries to figure out if an image is the original high-resolution, or if it’s been upscaled). The terrain looks significantly better, but I’m struggling with making runways and cityscapes look correct. Specifically, enforcing context (“there’s a brown rectangular patch next to what looks like roads and what looks like trees - is this a small field that’s the size of a large garden, or a house with a brown roof?”), and dealing with aliasing have been the main challenges. By aliasing, I mean that regular patterns like railways, farm plow lines, etc. are hard to recover correctly because many configurations at high resolution look exactly the same at low resolution. Here is an example:
TL: Original 0.5m resolution
TR: Upscaled 0.5m resolution
BL: Bilinear upsampled 0.5m resolution from 2m resolution
BR: Nearest upsampled 0.5m resolution from 2m resolution
Chances are that something like this will not affect gameplay at all (especially since I’ll probably only release at 1024x1024), unless the stripes end up being runway thresholds pointing the wrong direction or something funky like that. I’ll release a beta (it’ll still have a lot of artifacts in places, that’ll likely go away with training) tomorrow if some of you want to test it. This is what it looks like at the moment:
L: Original, nearest neighbour upsampled
😄 Original, bilinear upsampled
R: AiTiles 1m