11 hours ago, abrams_tank said:

If they could detect photogrammetry trees, they would have removed them in the first place. The reason why there are photogrammetry trees is because the photogrammetry technology they are using is mistakenly picking up trees as objects.

Artificial intelligence can do face recognition for quite some time. The differentiation between trees and buildings shouldn't be such a challenge.

16 hours ago, UAL4life said:

I don’t normally defend Asobo, but PG can’t differentiate between trees buildings or cars. PG just has 3D scan data…
 

Maybe in the future Ai will be able to pick apart PG but I doubt it will be soon because even the PG is very varied and fluid (hence the sometimes melted look)

There is zero reason they couldn't run the PG data through a ML algorithm to identify trees and remove them. They could look at both the spatial structure of the "trees" and also the textures themselves. Sure, it could pick up the random smokestack or light pole, but the overall look would be significantly better IMO. They are almost certainly doing this already to some extent to clean up the raw data coming in from the planes and drones. I don't know why they feel a weird ugly polygonal tree looks better than their beautifully-crafted autogen ones.

Duplicate

Edited December 4, 20241 yr by kiwikat

I was thinking the same, there should be a way to be rid of the PG trees, they serve no purpose.  If their AI can generally identify what is a house, what is a tree etc, one would think it’s not a stretch to identify trees in PG areas and replace them with autogen trees (which look generally great in FS24).

Since their ML algorithms already detect trees based on satellite images, maybe use that to automatically create exclusion zones that disables PG data around detected trees. Might be easier to implement instead of training ML on 3D data.

2 hours ago, megasus said:

Since their ML algorithms already detect trees based on satellite images

That's what they claim. But I found that their algorithm is pretty stupid. If there is no OSM Data to help it it misenterprets the images all the time.

Examples are for example the "Kitzbüheler Horn" in Austria, where even in 2024 the algorithm can't tell where it should place trees and where not. Another example is the Namibrand Naukluft park in Namibia, where hills that are in real life nothing but blocky rocks are covered with trees in the sim.

Or that it does not seem to know where the treeline is or that trees don't grow in the cold arctic climates.

Edited December 4, 20241 yr by Farlis

On 12/3/2024 at 2:46 PM, kiwikat said:

They 100% could if they wanted to. This would not be hard for a machine learning model.

I definitely prefer '24 with PG turned off if flying around cities- they look like the rain forest otherwise. 

So a photogrammetry expert answered why taking out the trees is actually very hard and expensive to do:

Quote

I can’t speak with great confidence for MSFS, but what you’re talking about is certainly possible, but highly impractical.

For one, it’s insanely expensive to build, train and validate a model like that. Not to mention time consuming. You would either need to buy an off the self solution, which doesn’t really exist or make doing that an entire division in your company and then sell it as a product or offer some kind of expensive B2B type service using it to justify the investment in it.

Second, it’s certainly possible to do this at a small scale (in the hundreds of models), but based on my experience, you’re looking at an hour or more as soon as you do anything that boils down to “AI, do these things to my 3D model”. Now our AI guys could just suck, but I don’t think they do. So at the scale they’re operating at with these models, you’re probably talking about years to run them all, if not decades (I don’t actually know the amount of 3D models they have, but you could be generous and say an hour per model to do what you’re suggesting).

Last, it’s just not strictly needed for this type of game. The models only really need to look good from far away, which they mostly do. They go in and handcraft anything they deem important to look good up close, and that’s good enough for most cases.

This person, if they are a photogrammetry expert, is claiming it would take years or even decades to remove all the photogrammetry trees.

And you can read about this person's credentials and work experience here:

Quote

Yeah. We do photogrammetry as part of the services where I work. Even with heavy processing and high-quality images from a pre-programmed drone that takes the optimum images, the quality can be hit or miss. We have a version that we do cloud-based rendering on, so we can use the full-quality model and that usually ends up being good, but when we start trying to optimize it to render on the client side, it gets a little iffy. The "good" photogrammetry model tends to be massive, far too massive to render on the client (like 2+GB for larger buildings).

@kiwikat, sometimes it's never that easy or cheap to do, as you think it is. Sometimes you think a matter is pretty trivial technically, but it turns out that it's actually hard to do, and/or very expensive to do, making it not feasible.

Like I said, if the MSFS team could remove and replace the photogrammetry trees easily, they would have done that a long time ago.

 

Edited December 6, 20241 yr by abrams_tank

I'm focusing on removing photogrammetry (PG) trees and houses, as they often overlap. Since the houses are usually distorted and unrecognizable, I'm leaving only the main buildings and high-rises intact. This streamlined approach allows for much faster processing.

On 12/6/2024 at 1:34 AM, BijanStudio said:

I'm focusing on removing photogrammetry (PG) trees and houses, as they often overlap. Since the houses are usually distorted and unrecognizable, I'm leaving only the main buildings and high-rises intact. This streamlined approach allows for much faster processing.

This sounds like a realistic approach considering the overlaps. Small buildings and homes are pretty well represented by autogen, compared to larger businesses and high rises. Have you been able to scale this up to process large areas? This is the kind of stuff I was hoping Microsobo would do with their teams of data scientists and near limitless compute in Azure.

Thanks for your efforts! I'm off to purchase your 2024 product now to support your development.

14 hours ago, kiwikat said:

This sounds like a realistic approach considering the overlaps. Small buildings and homes are pretty well represented by autogen, compared to larger businesses and high rises. Have you been able to scale this up to process large areas? This is the kind of stuff I was hoping Microsobo would do with their teams of data scientists and near limitless compute in Azure.

Thanks for your efforts! I'm off to purchase your 2024 product now to support your development.

Thanks so much for the support 

I could be wrong but I doubt Asobo will have the technology for this. 

For now I'm doing 2 or 3 cities a week. I finished Vancouver, part of Los Angeles, Berlin, and few other ones. 

 

1 hour ago, BijanStudio said:

Thanks so much for the support 

I could be wrong but I doubt Asobo will have the technology for this. 

For now I'm doing 2 or 3 cities a week. I finished Vancouver, part of Los Angeles, Berlin, and few other ones. 

 

Whatever you do, keep those exclusion files seperate so that we can switch them off, when we don't want them.

I don't mind the PG trees (I mind the autogen trees that grow on top of them), and I don't want to lose a single house because of an exclusion. After all it could be one that I know from the real life. 😉

Edited December 8, 20241 yr by Farlis

1 hour ago, BijanStudio said:

Thanks so much for the support 

I could be wrong but I doubt Asobo will have the technology for this. 

For now I'm doing 2 or 3 cities a week. I finished Vancouver, part of Los Angeles, Berlin, and few other ones. 

 

Microsoft has the technology to do this with its Microsoft CoPilot tecnologies. As a matter of fact any Strong AI (generative AI) from any of the other companies (OpenAI, Google, Oracle, etc) can easily be trained to do this. Heck, it's better suited to a small, weak AI (specific task) model (see ArcGis) to do exactly what you are spending your valuable time doing.

The irony is that as much as Microsoft touts its AI technology in the market, it doesn't showcase any of it in MSFS 2024.

Edited December 8, 20241 yr by Mike T

7 hours ago, Mike T said:

Heck, it's better suited to a small, weak AI (specific task) model (see ArcGis) to do exactly what you are spending your valuable time doing.

The irony is that as much as Microsoft touts its AI technology in the market, it doesn't showcase any of it in MSFS 2024.

Hey that's what I thought too, but I am only a data engineer and GIS admin. Admittedly I don't know the specifics of working with flightsim scenery but from the data side this doesn't sound like an unsolvable problem. It might be challenging and require a lot of ML skills that I certainly don't have, but nothing about this sounds unsolvable given what I've seen my team do in the past and me having slung a healthy amount of geospatial python.

At the very least I'm super grateful for folks like Bijan chipping away at it! 😎

Edited December 8, 20241 yr by kiwikat

Photogrammetry is part of msfs; and exactly where the trees are; also they are the exact height; taking it away is like breaking the sim.

46 minutes ago, Marc22 said:

Photogrammetry is part of msfs; and exactly where the trees are; also they are the exact height; taking it away is like breaking the sim.

Exactly. On flightsim.to you will find a few add ons that clean up the PG. But not by removing the PG trees, but by totally removing the autogen trees in the PG areas.

Looks great, and is my prefered solution.