I had a look at some more places and compared them with OSM.

We already discussed this location a while ago. Here it seems that areas where OSM has data are also more accurate in the sim, whereas where there’s no data in OSM the buildings in the sim seem to be placed in a rather random fashion.

Here, similar to the Tokyo shot, spaces that are empty in OSM are populated in the sim, but in this case much more accurately than in Tokyo.

Here OSM and sim overall seem to match. Also more complex footprints such as the tower blocks at the bottom appear to be identical; however, there are also some discrepancies (marked red). While in Tokyo some other data or the AI seems to be used to fill the blanks left by OSM, here we have examples of buildings that are present in OSM but not in the sim. How does that make sense? Or could it be that the sim’s algorithms struggle with certain shapes (e.g. towers on top of a pedestal building like the one on the left). Or does it mean that our theory (i.e. OSM as source where there‘s no Azure footprints) is wrong altogether?

Here too it seems that those buildings that are in OSM are placed more accurately in the sim. What can be ruled out I guess is that the sim uses OSM data for building hights. There are just too many deviations.

But here too do we have some discrepancies between OSM and the sim as can be seen here. The tower is missing and the building in the foreground has a different footprint (again, this could be because of its shape (pedestal building)). This comparison also shows how accurately roof colours are adopted.

Here two places where we don’t have footprints in OSM:

Again, the buildings seem to be placed rather accurately. In Tout Lake we can once more see how roof colours are depicted accurately.

What does this all mean now? Is OSM one of the sources or not? Or is it a coincidence because OSM partly has the same source data as the sim (e.g. Microsoft’s footprints). Honestly, I have no idea. This comparison made the whole thing only more complicated for me 😅.

Edited May 3, 20206 yr by Shack95

I think the answer is in episode one. AI, relies on biomes and data sources to build the world. So it's possible that OSM is one source, but not the only source, and the AI could be picking up data where it finds it, from a variety of data sources, including OSM, including bing, etc...

3 hours ago, Dominique_K said:

I am afraid I don't belong to the 'all'. The AI is of course responsible or rather how it has been programmed. It needs to be trained to interpret imagery, it is a long process of trials and errors and what the picture show is its young age and lack of supervision. Training could mean, for instance, having now a constraint telling no building higher than two storeys in a well defined Bangkok area. Your second arrow from the left is, in real Bangkok, a perimeter wall. Other arrows are low buildings. Same could be said of the buildings along the river.

As a generic  user I am only interested by the end result not the technicalities explaining me that I should expect it to be  wrong.  Asobo said they were serious about delivering the best, and I think they are, they'd better have a systematic review of high visibility places like Bangkok.  

 

What I am showing is simply that the OSM data is most likely, one of the data sources they are using to build the world infrastructure, the same way we do it for the other simulator (i.e: transform OSM footprint into 3D infrastructure based on rules, filters, conditions etc.)

When we create global scenery using a very similar process, and with OSM data, as I did for the other sim, I think we can identify the same process in another simulator no doubt 🙂

1 hour ago, Shack95 said:

What does this all mean now? Is OSM one of the sources or not? Or is it a coincidence because OSM partly has the same source data as the sim (e.g. Microsoft’s footprints). Honestly, I have no idea. This comparison made the whole thing only more complicated for me 😅.

Wow, thanks for this detailed analyzes 🙂

Yes I think that it is almost clear now that OSM is one of the data sources.

Also, I believe that for the US and Canada, the millions of footprints Microsoft scanned using its AI and released a couple of years ago with a White Paper describing how it was done, they are no doubt also used.

Now "the little house in the prairie" 🙂 you show in the last example, is it in the US? If it is, then it could be from the US in-house scanning I just mentioned.

And who knows, Bing might have its own footprint data.

Now as I mentioned before, if data is not available for specific areas, we apply purely procedural methods (ex: neighborhood of tokyo). That means without data, we write rules to plant on each side of a road segment, or inside the lots created by the intersections, specific infrastructure based on different parameters. This is almost a random or we can call it a plausible result to fill empty spaces. The 3D infrastructure is usually ready made (library objects) and of course they don't match the aerial imagery infrastructure at all.

I think to have an idea on how all this is built, one must explore the documentation of two procedural applications for existing flight simulators: 

SceneProc for FSX: https://www.scenerydesign.org/scenproc/

World2XPlane (I placed the documentation on my Google drive) : https://drive.google.com/file/d/1TikGob_tmY8-3zQAX-shr5y-xqG7-fL2/view?usp=sharing 

8 minutes ago, Claviateur said:

What I am showing is simply that the OSM data is most likely, one of the data sources they are using to build the world infrastructure, the same way we do it for the other simulator (i.e: transform OSM footprint into 3D infrastructure based on rules, filters, conditions etc.)

When we create global scenery using a very similar process, and with OSM data, as I did for the other sim, I think we can identify the same process in another simulator no doubt 🙂

I understand and appreciate the technical background that you bring to the debate ! A pity that you are not the in the alpha.  But I wish to place it on another ground. MS decided a year ago to engage the community and I intent to play by their rules. When something is wrong (or I think it is wrong) I tell it as I see it. Asobo promised a next gen sim not another OSM-driven XP ( appelons un chat, un chat 😉). They need to look closely at their automation process.

ANd thanks @Shack95 for his post up above !

1 hour ago, azulkb said:

I think the answer is in episode one. AI, relies on biomes and data sources to build the world. So it's possible that OSM is one source, but not the only source, and the AI could be picking up data where it finds it, from a variety of data sources, including OSM, including bing, etc...

For sure, the Episode 1 explains perfectly the processes and we can see clearly that not only one data source is used but multiple and an augmentation is done with extra data (from Azure AI) as well as procedural generation. 

 

Edited May 3, 20206 yr by Claviateur

23 minutes ago, Dominique_K said:

I understand and appreciate the technical background that you bring to the debate ! A pity that you are not the in the alpha.  But I wish to place it on another ground. MS decided a year ago to engage the community and I intent to play by their rules. When something is wrong (or I think it is wrong) I tell it as I see it. Asobo promised a next gen sim not another OSM-driven XP ( appelons un chat, un chat 😉). They need to look closely at their automation process.

ANd thanks @Shack95 for his post up above !

If I am in the Alpha I would stop working on my actual scenery so probably it's a good thing in my case 🙂 

Yes I am a 100% with you on giving feedback on things that do not seem right but now that we understand how the world is built and as it was reminded by the post above, Episode 1 gives some hints as for the techniques used to build the world.

Yes the AI is scanning and fetching extra data as they said in Episode 1 but as far as we can see, this scanning process does not seem (yet) world wide when it comes to infrastructure. Probably for vegetation zones it did so... Hence the OSM data being used + purely procedural processes.

Now that we know what techniques are applied, to refine the world using the actual techniques we see, the segmentation of rules and filters and all the conditions and sub-sub-sub conditions one can write in the configurations of the data or procedural scenery generation processes can have an impact on the result. But then how diverse is the visual material to represent the different variation of facades and roofs around the world is another aspect of it.

You know, if Microsoft will continue to feed the OSM with its data as it did with the Azure footprints for US and Canada, OSM will be a very valuable source of data. The problem with OSM is when data is manually added by people with different levels of knowledge and patience 🙂

 

Edited May 3, 20206 yr by Claviateur

@Claviateur Like pointed above by the others, it's not that simple. They are not simply using OSM data on the sim.

There's for sure, from my point of view, A.I. scanning involved all around the Globe, they've already used it to scan the trees and general vegetation, why wouldn't they use the data for the footprints of the buildings? Look the color of the roofs, for example, they are (almost) always matching the photoimagery.

My point here is, the quality of the "autogen" will largely depends on the quality of the imagery. Let's say the A.I. isn't intelligent enough right now , it needs more "practice", and sometimes it doesn't get right things from the poor quality imagery, things like agglutinating multiple buildings into one or not spotting at all some buildings. My hope here is the photoimagery and the A.I. will improve within due time and we can see a better/more accurate depiction of the cities.

We have to look outside Europe (where OSM isn't so good). If they weren't scanning the globe for the buildings, we would have flat photimagery basically everywhere outside Europe or the areas covered by photogrammetry.

Here's an example of the City of Guarulhos, near the city of São Paulo, the roofs are color matching, also there's no footprint on Bing or OSM for the area:

1 minute ago, ca_metal said:

@Claviateur Like pointed above by the others, it's not that simple. They are not simply using OSM data on the sim.

There's for sure, from my point of view, A.I. scanning involved all around the Globe, they've already used it to scan the trees and general vegetation, why wouldn't they use the data for the footprints of the buildings? Look the color of the roofs, for example, they are (almost) always matching the photoimagery.

My point here is, the quality of the "autogen" will largely depends on the quality of the imagery. Let's say the A.I. isn't intelligent enough right now , it needs more "practice", and sometimes it doesn't get right things from the poor quality imagery, things like agglutinating multiple buildings into one or not spotting at all some buildings. My hope here is the photoimagery and the A.I. will improve within due time and we can see a better/more accurate depiction of the cities.

We have to look outside Europe (where OSM isn't so good). If they weren't scanning the globe for the buildings, we would have flat photimagery basically everywhere outside Europe or the areas covered by photogrammetry.

Here's an example of the City of Guarulhos, near the city of São Paulo, the roofs are color matching, also there's no footprint on Bing or OSM for the area:

Yes you are right about the example you show. But it is not the case for other parts of the planet we saw in the screenshots...  

Scanning could be still ongoing for the rest of the planet who knows. And it could be that in the meantime, other data is used to populate the world while waiting for the scanning... Or a mix of both...

The examples we were analyzing were certainly not the result of any scanning as in the case of your examples. Perhaps in the case of the screenshots we were debating, the particular sector of the city is OSM while other sectors are from a scanned data, who knows...

But then even without scanning, or OSM or any data source, we won't have flat photo imagery, because as I mentioned previously and Episode 1 mentions as well, there are procedural techniques to fill the gaps.

13 minutes ago, Claviateur said:

Yes you are right about the example you show. But it is not the case for other parts of the planet we saw in the screenshots...  

Scanning could be still ongoing for the rest of the planet who knows. And it could be that in the meantime, other data is used to populate the world while waiting for the scanning... Or a mix of both...

The examples we were analyzing were certainly not the result of any scanning as in the case of your examples. Perhaps in the case of the screenshots we were debating, the particular sector of the city is OSM while other sectors are from a scanned data, who knows...

But then even without scanning, or OSM or any data source, we won't have flat photo imagery, because as I mentioned previously and Episode 1 mentions as well, there are procedural techniques to fill the gaps.

My point is, just because we are spotting mismatches on the scenery, doesn't necessarily mean they didn't scan the area. It could just mean the A.I. didn't get it right or the imagery was of too poor quality that it filled it up with procedural data.

Look again at that Bangkok shot and will notice that the roofs are color matching with the imagery. Those buildings with the real colors are probably collected within an scan. Some buildings were agglutinated, but still have the right colors.

I've been assessing all the screenshots and there are buildings in the middle of nowhere, accurately shaped, positioned and their roofs have the right color. Why would they scan the countryside of India or Chile and don't do the same for a major city like Bangkok?

Finally, it's still Alpha, we have almost a decade to it evolving. I'm sure (and they've already said it) the scenery will be constantly updated/improved. We will probably see better data and more regional facades and objects being added in the future.

 

2 hours ago, Claviateur said:

Also, I believe that for the US and Canada, the millions of footprints Microsoft scanned using its AI and released a couple of years ago with a White Paper describing how it was done, they are no doubt also used.

Now "the little house in the prairie" 🙂 you show in the last example, is it in the US? If it is, then it could be from the US in-house scanning I just mentioned.

Yes, the „little house“ is in the US (Trout Lake, Washington) and it makes sense of course that they use their in-house data. But didn‘t they say that these Azure footprints are also in OSM? Because in this particular case they aren‘t.

1 hour ago, ca_metal said:

We have to look outside Europe (where OSM isn't so good). If they weren't scanning the globe for the buildings, we would have flat photimagery basically everywhere outside Europe or the areas covered by photogrammetry.

Here's an example of the City of Guarulhos, near the city of São Paulo, the roofs are color matching, also there's no footprint on Bing or OSM for the area:

The same can be seen in the images I posted above. Even when there‘s no OSM footprints, the placement of buildings in the sim is pretty exact. Also the roof colours match Bing imagery. This indeed speaks in favour of imagery scans (as described in the discovery episode).

Still, it seems those buildings that are also in OSM are placed more accurately than the others which would support the theory that OSM is one of the sources as well.

Edited May 3, 20206 yr by Shack95

3 minutes ago, Shack95 said:

 

Yes, the „little house“ is in the US (Trout Lake, Washington) and it makes sense of course that they use their in-house data. But didn‘t they say that these Azure footprints are  also in OSM?

The same can be seen in the images I posted above. Even when there‘s no OSM footprints, the placement of buildings in the sim is pretty exact. Also the roof colours match Bing imagery. This indeed speaks in favour of imagery scans (as described in the discovery episode).

Still, it seems those buildings that are also in OSM are placed more accurately than the others which would support the theory that OSM is one of the sources as well.

Yes and as Ca_Metal mentioned above, I suppose the roofs, if accurately represented, could be a seperate type of data they fetched from Bing imagery scans, is mixed with OSM when required or with in-house data when available.

 

On 5/3/2020 at 10:26 AM, Claviateur said:

And the footprints in this area are traced by people who decided to draw wide infrastructure for some reason, instead of the smaller ones we see in aerial imagery. Probably because they wanted to go fast who knows... 🙂  Contributors to the OSM database sometimes trace footprints in a more precise way than others and some take the time to add tags while others just trace the shape... 

Yes, so this....  Asobo should err on the side of smaller buildings and presume that footprints over a certain size are "data-errors", and fill the footprint with smaller 1 storey (75%) and  2 storey (25%) buildings... eg, those walls that have been turned into massive walls of apartment blocks in the Bangkok shot.

Edited May 4, 20206 yr by MatthewS

4 minutes ago, MatthewS said:

Yes, so this....  Asobo should err on the side of smaller buildings and presume that footprints over a certain size are "data-errors", and fill the footprint with smaller 1 storey (75%) and  2 storey (25%) buildings... eg, those walls that have been turned into massive walls of apartment blocks in the Bangkok shot.

Divisions of footprints is possible, Asobo have all the skills to do that. So I am sure they can do it if it's on their priority list or if MS management let them put it on their priority list... 

But then there should be an additional set of conditions to make this process convincing. For example, to know if it's a bundle of infrastructure drawn manually as one footprint or if it's rather an actual wide warehouse in reality or a wide hotel or any other wide platform, they need additional infos like the type tag entered in the footprint.

Now, if the type is not tagged in the footprint by the person who traced it, they will need to identify the sector (industrial, residential etc.) This is also, sometimes available in OSM. Otherwise, the division will be as random or arbitrary as the manual grouping... And people will complain about hotels and warehouses become shops and houses.

I personally thought initially that the grouping is done on purpose or through automated processes in the procedural generation of infrastructure to simplify and reduce the shapes for the sake of performance... (i.e: in the other sim the facades eat a lot of fps especially if they have multi textures. Yet maybe this is not relevant to the MSFS engine)...

But after I saw in the discussion here that this is from the original OSM footprints, I went to inspect the results of my own OSM driven scenery in the other simulator. Believe it or not, there as well, I discovered so many wide footprints that were drawn to group many smaller infrastructure. Of course, when I created my procedural scenery (for 4 geo tiles only),also from raw OSM, I never went to check every street and sector and compared it to aerial imagery... Now I did 😕  

People who trace them either use a zoom out level that gives the impression of a wider infrastructure or they are in a hurry to fill the sector...

So I suppose Asobo won't inspect the whole planet for this type of automation. However, yes, if they can set logical conditions based on additional infos as mentioned above, they can divide footprints...

 

Do we know, for sure, that the OSM data is 100% accurate?