If you give AI a bunch of socks to build houses, all the houses will look like socks.
The building textures and shapes and sizes are just all wrong.
How is it that the AI can place 10,000s of 1 story bungalows in a city full of 3 story tall buildings?
And how is it that the door types, and windows and roofing tiles and chimney’s don’t look anything like they do in real life?
It just seems like the developers fed the AI with generic textures not really relevant to the area that much.
It’s rare to find a house in many UK Cities that actually look like houses you’d see anywhere in UK cities (that’s what I can say because live in the UK)
The autogen houses aren’t designed to look just like their real world counterparts. They just have to give you the impression of a house, not “that” house.
PG is designed to fulfil that need.
If you are concerned about what kind of roofing tiles, and chimneys they show, then I suspect you are flying too low, and I have a number for you.
^This so much. Autogen and even PG were designed to give a passable, relatively accurate rendition of structures and cityscapes at altitude.
It was never intended to hold up at close Hollywood stunt flying through the open horse barn ranges. Although you can see in many PG instances, it was good enough to capture large advertisements and store signs legibly. Heck, you’d be surprised at how well PG was captured in stock cities like the Las Vegas strip (not payware). Take a pass down there at very low altitude.
Recreation of a 100 percent fidelity Living World isn’t possible to be delivered either using existing data or delivery rendition mechanisms today. In the future, sure, let’s be optimistic about it.
AI would simply try to interpolate what it thinks it’s there. AI/ML/Models are only as good as the data the algorithms are trained on. Big Data (which covers the major FS sub-services: Live WX, Bing Maps, PG, Autogen, Live Traffic) have gaps and flaws in them. Those gaps and flaws will be reflected in the AI/ML/Model’s performance. You can’t train out errors without a completely pristine data set, which doesn’t exist in any of those domains.
This is a SaaS application delivering a more than 80/20 product set. That last 15-20 percent is an incredibly hard threshold to overcome, AI or not.
Gosh, some of you need to go back to P3D or FSX and see autogen buildings in those games lol (or hell, even Xplane). I’m often the first to point out that MSFS looks more like City Skylines than the real world, but come on guys, you just aren’t going to get accurate procedural building generation across an entire planet.
AFAIK, the AI relies on zoning data, images, and the 3d models it’s been given. If somewhere is zoned industrial, and there is big white square in the ground texture, the AI plonks down a big white factory building. I’m sure it’s a little more complicated than that, but expecting the building type to be accurate, let alone matching the roof style, is a bit of an ask, and I think you’ll be a bit disappointed if you expect it to be much better in MSFS2024.
I would agree that they could perhaps add more generic models to at least flesh out the variety and give some regions buildings that are more like the local architecture. Perhaps sets of procedural buildings per country, with maybe a few sets for large countries with multiple climates? But beyond that I wouldn’t expect much more.
Realistically the service would never switch to a competitor’s satellite imagery (or more accurately a cleaned up version - Google does the exact same QC changes as described by Martial in past Q&As). The reason they can provide Map Streaming to begin with is leveraging an existing subsidiary service that rolls up to MS, otherwise the cost would have to get passed on to the customer.
It seems like no one really listens to the Q&As. Jorg, Seb and Martial have many times already described that data sets vary in quality and accuracy/precision from a limited number of satellite imagery providers. And none of them can control the weather, especially over cloud-prone areas. Add to that, nation-state limits over image coverage and detail - and it’s not surprising that we have a patchwork quilt of terrain imagery that vary in quality, coverage, detail. This is the same situation any other company trying to recreate a Living World will encounter. MS just happens to be the lead firm to do so, so it’s easy to criticize when no one else is trying. Shrug.
This is an interesting topic though on how AI will impact video games in the future, and suspect MSFS someday as well. I watched a really interesting video earlier about how Rockstar is going to make 70% of all buildings enterable in GTA6. This is a huge jump from 47 in GTA5 to prob hundreds in the new version. It is going to be utilizing AI to create unique 3D elements (furniture, appliances, etc), so each interior appears different. Or at least how I understood it.
I could see something like this being used for MSFS for many of the buildings and objects already being generated. At minimum, lets hope they are improving the autogen engine so that we see more variety of objects, veg, etc…
Indeed. It takes a lot of workload off the devs. Although I think the term AI is perhaps getting a tad overused these days. A lot of what people call AI seems more like procedural generation than artificial intelligence.
Blackshark.AI has a lot of contracts with other companies over their use of the algorithm, so I am sure in 4 years it has progressed.
Hopefully, it will be able to differentiate different types of buildings. For example, I fly a lot in South Africa. If you fly over a squatter camp (an area where there are a lot of closely placed tin shacks and cargo containers) that ends up getting represented as a warehouse. If you fly over the townships, they are all small houses (I know my wifes mothers house was roughly the total size of just my living room, and my house is only 1200 sq feet) but every house has a wall around it for safety (typically with barbed wire at the top of the wall). Well that gets picked up as a large two story house. So you fly over a township, and it feels like you are flying over affluent neighborhoods when its quite the opposite in fact.
I don’t care if the houses are represented exactly as they are shaped, etc. But I do wish the general characteristics could be picked up and represented. If I am flying over a mall, I kind of want to know its a mall. If I am flying over a squatter camp, I kind of want to be able to tell. If I am flying over a township, I kind of want to know. It also helps greatly with VFR. For example, if I am flying from Soweto to the south east, and I go over a squatter camp, I know then the Golden Highway is coming up soon and Orange Farm and Springs (townships) are coming up soon. Right now, there is no distinction between the townships and perhaps some of the more affluent nieghborhoods in the Free State, and it makes it hard in these areas to navigate by landmarks.
Well, I will say this. There aren’;t always good sources for data that isn’t one of their competitors. I know at least for South Africa, using Google maps, I can drill down to the very yards in Soweto or Orange Farm and see cars parted in the yards, etc. Bing stops pretty far up. So yes, it would be hard using Bing’s current imagery for South Africa to determine buildings appropriately. But I do know Google has managed to get much better data.
Am I suggesting Microsoft make use of Google data? No. But Microsoft could invest in the area of say South Africa more. You might say “what could they get from that”? Well, Bing data ISN’T just used for MSFS. There is a whole lot of Uber, Door Dash type services, deliveries, etc. Every South African has a smart phone, and every one uses services on apps. It is almost as popular as it is in the US. But currently, all of those services drive off of Google because Bing data just isn’t reliable enough in the area.
So while I do get that Dem and such there may be issues getting good data, things that Bing can control, they don’t always invest enough in some areas. Thats a business decision, not a technical one.
I see here a huge thread based solely around a key misconception. Most of MSFS autogen is based on OSM zoning data bing acquired along the way.
Buildings in the middle of fields? That’s not AI. That’s a person tagging a field as an industrial structure on osm.
Low rise buildings in place of multi-story buildings? That’s osm structures lacking height data, feel free to fill it in…
To get a good example of what data does what see Chernobyl sarcophagus and powerplant. It is a unique place where DEM, OSM and photographic data are misaligned, so you can see which contributes what.
Orthophotography shows the sarcophagus before deployment, visible on texture to the west, but without a building
OSM has the sarcophagus object tagged in the deployed position, east of that. This generates a generic 3D building
DEM radar return read the deployed sarcophagus, so a hillock is clipping through the 3D building.
BLACKSHARK serves a limited role, supplementary to OSM zoning data. It matches rooftop color with the photo and in rare occasions attempts to place 3D buildings without osm blocks based on image recognition. So far it mostly fails at that, we have false positives on dunes and untagged villages with visible buildings without 3D
But one of the key features Jorg introduced at last years Flight Sim Expo was that FS2024 was going to introduce 2D-to-3D technology for buildings. Surely that would be going off the imagery in order to costruct what it thinks the building is.
We will see just how much they can actually do. So far it seems to me this was maybe tested in FS2020 but the main source for 3D buildings was definitely OSM
The thing is that google had a massive head start on the others, i dont know how many years, but it must be at least ten years. Thats a lot of time when it comes to refine, expand and perfect their map service. Bing is getting there though… but yes they are running behind.
