4 ms·

As with all these "this X doesn't exist" sites, the images look ok at first sight, but when you start to look closer you notice that some things aren't quite ri

by rob74 5y ago

As with all these "this X doesn't exist" sites, the images look ok at first sight, but when you start to look closer you notice that some things aren't quite right. In this case, it's mostly roads that don't seem to connect to anything...

ogig 5y ago
As someone spending time looking at satellite and orthographic photos, i agree, this looks weird. City grid looks like cloth texture, roads and rivers make no sense. In any case, it manages to slightly impress me.
- sdflhasjd 5y ago
  It's like it can't decide between a city on a grid or not, and instead ends up with a weird fusion of both - and like you said - looks like a woven fabric instead.
- dTal 5y ago
  >As someone spending time looking at satellite and orthographic photos That's presumably the same training the AI had :p
- warmwaffles 5y ago
  How does one get an orthographic photo that is not artificially made? I know what an orthographic camera does in 3D graphics, but how can you accomplish that in a real life?
  - detaro 5y ago
    they are "artificially made" by postprocessing aerial images. (You can actually do cameras that create orthographic images, but that's more for industrial vision etc since the frame covered obviously can't be larger than the lens, which doesn't really work for aerial photography, so if people talk about "ortophotos" in that sense they mean processed)
hajhatten 5y ago
I'm surprised you managed to notice that. The images are so low res.
- ajmurmann 5y ago
  The low resolution to me undermines the entire thing. On some of these images I couldn't even tell if there was even a city or just a barren landscape with a handful of warehouses
  - uoaei 5y ago
    All bow before the rough-approximation power of L2 loss functions!
notahacker 5y ago
Yep. I was quite impressed by how mine appeared to stick bridges in logical positions. Then I noticed one of them only got halfway across the river...
- brk 5y ago
  That's not a bug, it's just bureaucracy. Wait 6 months and refresh.
msdrigg 5y ago
I wonder if you could improve this GAN by taking a cropped part of each generated image and trying to discriminate against that part of the image as well. Because the problem seems to be in the details. The GAN is great at the global picture but the details are off. And this problem benefits from the fact that a picture of a section of a city should also be recognized as a picture of a city.

by rob74 5y ago

ogig 5y ago

As someone spending time looking at satellite and orthographic photos, i agree, this looks weird. City grid looks like cloth texture, roads and rivers make no sense. In any case, it manages to slightly impress me.

sdflhasjd 5y ago
It's like it can't decide between a city on a grid or not, and instead ends up with a weird fusion of both - and like you said - looks like a woven fabric instead.
dTal 5y ago
>As someone spending time looking at satellite and orthographic photos That's presumably the same training the AI had :p
warmwaffles 5y ago
How does one get an orthographic photo that is not artificially made? I know what an orthographic camera does in 3D graphics, but how can you accomplish that in a real life?
- detaro 5y ago
  they are "artificially made" by postprocessing aerial images. (You can actually do cameras that create orthographic images, but that's more for industrial vision etc since the frame covered obviously can't be larger than the lens, which doesn't really work for aerial photography, so if people talk about "ortophotos" in that sense they mean processed)

hajhatten 5y ago

I'm surprised you managed to notice that. The images are so low res.

ajmurmann 5y ago
The low resolution to me undermines the entire thing. On some of these images I couldn't even tell if there was even a city or just a barren landscape with a handful of warehouses
- uoaei 5y ago
  All bow before the rough-approximation power of L2 loss functions!

notahacker 5y ago

Yep. I was quite impressed by how mine appeared to stick bridges in logical positions. Then I noticed one of them only got halfway across the river...

brk 5y ago
That's not a bug, it's just bureaucracy. Wait 6 months and refresh.

msdrigg 5y ago

I wonder if you could improve this GAN by taking a cropped part of each generated image and trying to discriminate against that part of the image as well. Because the problem seems to be in the details. The GAN is great at the global picture but the details are off. And this problem benefits from the fact that a picture of a section of a city should also be recognized as a picture of a city.