Let's say I put some AI at work to scan the Bing aerial image of Larsen Bay, I do not know what zoom level the MS Azure AI uses but let's say the one I grabbed in the screenshot below...

I think and based on my observation of the screenshot the vegetation zone is more or less what I draw manually with red lines in this small portion

I don't think the detection is done one green patch at a time otherwise the data will be extreeeemly heavy for a small region. 

Edit: If it's detected one green patch at a time, I suppose grouping is done for processing matters and data heaviness.

So I presume the empty patches detected, and as seen in the screenshot, are no smaller than this (minus roads)

Edited June 3, 20206 yr by Claviateur

12 hours ago, Claviateur said:

So I could be mistaken but in MSFS the AI does not generate visual scenery. Asobo do from the available data with scripting they type themselves for the engine to understand and thus to render visual elements...

Thanks for your clarifications. I think that's how it works essentially. 

Sébastien Wloch explained it as follows:

Computers are today more used to manage the scale of the world […] thanks to Azure and running […] something like 20000 computers we’re able to tag every tree on the planet on the satellite photos in less than 7 days. I think it’s close to 3 trillion trees. It finds them all and you know where they are and it takes less than 7 days, so if the data is updated, we can get all the new trees in 7 days. That’s more what we use AI for, it’s machine learning AI. It can do stuff like this. I would say it’s not highly smart, it’s not like a human, but it’s very fast because you can have 20000 of it and it’s very fast at doing stuff like this, finding houses, finding what type of rooftop a house is […] and maybe in the future other stuff like [unintelligible], anything we wanna find on the planet. And also QA[?], it’s finding clouds right. Sometimes satellite data is dirty, there’s clouds, there’s areas which have been blurred for whatever security reasons or stuff like this. It can find data like this and tag it and then we can work with it. And that’s the way we use not only just AI, because if you only had AI on your computer [unintelligible], it would take 20000 times one week, it’s AI plus Azure because Azure can give us 20000 computers. That’s why it all comes together. (transcribed from AVSIM interview, 1 October 2019, 24:49-26:20)

So basically the AI (that's where Blackshark.ai comes in I guess) extracts the required data from satellite imagery and Azure is the infrastructure they're using to do this in an efficient manner. That's in line with what the guy from Bongfish said in that Youtube video posted a few weeks ago.

As to trees, Wloch's wording implies that they're tagging individual trees and not tree zones. But of course, tagging "every tree on the planet" could just be a more spectacular way of putting it. Or, as you said, they then group the trees to reduce data size. Who knows... 

Another point that has been discussed is whether the AI is used to remove clouds from the imagery. Apparently, the AI just detects and tags clouds, it does not retouch the imagery. 

What‘s very promising I think is that in the future the AI could possibly be used to find „anything we wanna find on the planet“. I wonder how far into the future they are right now.

Edited June 4, 20206 yr by Shack95

1 hour ago, Shack95 said:

What‘s very promising I think is that in the future the AI might be used to find  „anything we wanna find on the planet“. I wonder how far into the future they are right now.

 

Indeed. The interpretation of aerial maps (drones, aircraft, satellites) has huge commercial stakes and large companies put tons of money in  new AI applications. Count on MS to be in the game (pun not intended or ...maybe it should be intended). 

One cannot sum up AI to a simple pattern recognition. it has predicting and self-learning abilities. See for instance the fascinating mapping of the road networks in Thailand and Indonesia  https://tech.fb.com/ai-is-supercharging-the-creation-of-maps-around-the-world/. Same for the vegetation. Tree recognition is not for the simmer's satisfaction. It has a lot of commercial use. 

It is an evolving field and I surmise that, if MS/Asobo are true to their 10-year commitment, some FS sceneries will be in 2030 considerably better in 2030 than in 2020,  for two converging reasons : the profusion of new imagery (satellites are cheap to build and launch now and we are just at the dawn of the use of drones for commercial purpose) and  because of a much more sophisticated AI.

I don't know the state of advancement at MS and Blackshark.ai and I suppose all that is very proprietary and we will not see an EP anytime soon. But I sure know that we haven't seen nothing yet 😉 ! 

PS I am no specialist but I find all that fascinating. Plenty to read on the Internet.

Edited June 4, 20206 yr by Dominique_K

15 hours ago, jlund said:

It's not an excuse, but the ai will NEVER be able to get every place in the world correct. We'll all come over places that we know well, and will probably see things that isn't like in real life, BUT for people who doesn't know the excact place, it doesn't matter. I have no idea on how ie.Larsen Bay in Alaska looks, so I'll probably don't care if the vegatation isn't 100% correct. 

When I go out to explore in Flightsim I always check how the real places look like, either before or after my flight.

And nothing bugs me more than the wrong vegetation.

If anything, this new sim and our discussions have the merit to raise the question of the importance of vegetation in the flight simulation immersion. At least for simmers flying below 12Kf.

I am convinced that the AI technology will bring us close to a faithful representation of the real world in this regard : placement, size, varieties and, yes, also the seasonal flowering. We cannot think that yesterday tech (including the TE type) with their limitations does reflect what we can get now and tomorrow with the AI.

4 hours ago, Shack95 said:

As to trees, Wloch's wording implies that they're tagging individual trees and not tree zones. But of course, tagging "every tree on the planet" could just be a more spectacular way of putting it. Or, as you said, they then group the trees to reduce data size. Who knows... 

Thanks for the infos, I missed the detail about individual tree tagging. This is impressive and huuuuuuge! I wonder how they handle this data or, perhaps, as I mentioned previously, they could group them into zones.

But point based database for the world forests is a huuuuuge data to work with and I can't wait to see how it was translated into 3D assets.

1 hour ago, Dominique_K said:

I am convinced that the AI technology will bring us close to a faithful representation of the real world in this regard : placement, size, varieties and, yes, also the seasonal flowering. We cannot think that yesterday tech (including the TE type) with their limitations does reflect what we can get now and tomorrow with the AI.

Agreed. For me the true revolution in this flight sim is its use of such technologies. As the AI and TTS improve; as Bing, Navblue and whatnot update their data, the sim is getting better and better. And the good thing is, these resources have nothing to do with MSFS. This allows continuous refinement at relatively low costs.

Of course there are also downsides, e.g. that the devs can‘t decide which areas have photogrammetry or that ATC doesn’t know how to pronounce JSTRM correctly, but all things considered I think this is a massive leap forward.

Edited June 4, 20206 yr by Shack95

1 hour ago, Dominique_K said:

If anything, this new sim and our discussions have the merit to raise the question of the importance of vegetation in the flight simulation immersion. At least for simmers flying below 12Kf.

I am convinced that the AI technology will bring us close to a faithful representation of the real world in this regard : placement, size, varieties and, yes, also the seasonal flowering. We cannot think that yesterday tech (including the TE type) with their limitations does reflect what we can get now and tomorrow with the AI.

The AI today is capable of doing amazing things but the question is always per project / budget and pragmatism I suppose.

For MSFS, as Shack95 clarified, the Azure AI is building a database by fetching data from scanning and putting it on the desk of Asobo to work with. The clouds for example are detected and tagged and Asobo deal with the retouch in some way or another (manual or Photoshop scripting who knows).

So as I initially thought for MSFS, the Azure AI is a data fetcher and this is a great thing already. 

And I suppose this makes full sense because creating a procedural world from traditional scripting (or node base scripts as in modern engines) using data from any source including AI scanning, is less complex in my opinion than having an AI that builds a scenery visually, retouch imagery etc.

For now at least and the upcoming few years, I think it's a cheaper and more logical for a video game (MSFS or any other MS game project) to continue with traditional development methods while using machine learning for data augmentation as Asobo mentioned...

But will the AI one day be a video game creator (code and visual)... No doubt!

 

1 hour ago, Shack95 said:

or that ATC doesn’t know how to pronounce JSTRM correctly

This could probably be solved. Assuming the exceptions are a small percentage, the program could look up identifiers such as these in a list of exceptions that would then say "JSTRM=Jetsream" or similar (don't want to be too technical). All other ones are given to the TTS as they are. The ATC text could then still display JSTRM but say Jetstream.

Dresden, Germany

Alicante, Spain

Mont Cenis, France

@Shack95 Alright for the longest time I was trying to figure out how Shack95 is finding all these freaking places? It just came to me, that he is an AI, Azure, whatever you want to call it and NOT a human being, IBM Watson meet Shack95 🙂 Dude you are amazing, as well as the other people that find these places, but mainly you, Kudos! Hashtag, Shack95  is a Robot 🙂 And waht's really funny is I am old and don't even know how to use The Twitter 🙂

Southeast of Cañas, Costa Rica

@Shack95 you are the best!!