Is there an easy way to see that this problem is complicated enough to require a neural network? I'm a noob in graphics, but this problem seems like it shouldn't be that difficult to require a neural net.
It does look like it would be doable without neural networks. My guess (having read their paper but not worked in the domain) is that it would be doable but requires a lot of iterations, tweaking parameters and dealing with corner cases. That would give you something running almost instantaneously but it would take a year to develop with tight feedback from domain experts.
Or, you can throw a (relatively standard for that kind of task) neural network at the problem, have a prototype at the end of the week and something solid by the end of the month.
In those conditions it makes sense to go with the neural network based solution (however, I do hope that having this result will push people to work on a more traditional approach that is competitive with the neural network based one).
The chances of the problem not having being already tackled with deterministic algorithms is the thinnest - and in fact, yes, there do exist tools, as expected.
It has but I am hoping that a paper accomplishing the task with machine learning will move the goalpost (it has happened in other areas of computer vision).
I don't think I understand your expression there: do you mean that "the problem of producing relief bitmaps from elevation data could lead to new achievements in our knowledge relevant to Neural Nets"? I am not sure why.
If you mean, more consistently with your original parent post, "lead to improvements in the past traditional techniques towards the same goal" - I would say that I don't see much space for leaping better results than those already possible.
I mean that results that would have previously been considered good enough for classical methods could now be considered improvable leading to renewed work into application of traditional methods to this subject.
We have been doing it "since forever" without Neural Networks.
Check Stefan "der-stefan" Erhardt's instructions to create your own OpenTopoMap service and map tiles from iron: you download elevation data and process them into shaded elevations - https://github.com/der-stefan/OpenTopoMap
I think a lot of it has to do with the fact that a 2D map is going to have roads and valleys and big green spaces in the mountainsides and peaks. The neural net identifies that villages and curvy roads represent valley floors and interpolate where the mountain slopes are.
I think it's using the contour lines on the map, not the villages and roads. The paper mentions the training data is contour maps and relief maps of the same area.
There's a nice example on that page with an overlayed relief shaded map and the original. Can you explain in English step by step how you would turn one into the other manually using something like photoshop?
Use the partial derivatives of the topo height to calculate a normal. dot that normal against your fake sun angle to shade your terrain.
it seems though that the magic in this is in throwing away specific details and changing the sun angle when it is aesthetically pleasing which makes sense.
on the surface it looks like you should be able to use the heightfield and standard filters to get a reasonable result but I'll take their word for it that real world data will prove more difficult.
If you read the text of the site or the paper: it needs a digital elevation model (i.e. a height map) to create the shaded relief. It does not ingest graphical maps and does the relief, it needs a discretized map of height data, such as lidar (which can be downloaded for any place on earth with 1m resolution).
Or, you can throw a (relatively standard for that kind of task) neural network at the problem, have a prototype at the end of the week and something solid by the end of the month.
In those conditions it makes sense to go with the neural network based solution (however, I do hope that having this result will push people to work on a more traditional approach that is competitive with the neural network based one).