Unlimited Detail, octree traversals

[Laa-Yosh]

And quite ugly, too.

I wonder when they'll realize that this video is only going to demonstrate just how much better current games look than their stuff. If they want to gain mindshare, they'd need to hire at least one good artist and put together a demo that's only 1-2 minutes short but totally impressive. Most people don't understand a bit of the tech but they dig the UE and id and Crytek demos because they have quality content. This 10-minute bashing of clearly superior looking games isn't gonna get them anywhere.

Not that I'd see much use of their tech anyway, though, it's most likely a dead end.

 

[fellix]

Well, as Carmack already tweeted, this technique is still way off its practical application. But I would be rather happy for a simple fly-by tech demo.

 

[Xenus]

Well their art improved over last year that's about it though.

 

[Grall]

I wonder how the hell we're going to animate the equivalent of a hundred vertices per square millimeter of game geometry. Sounds to me like the GPU's shader units would just roll over and die from that.

 

[Platon]

A game like Final Fantasy XIII would be a good candidate then; almost all of the environments are static and non-interactive. One question, is it possible to combine a voxel based engine with a polygon one? I mean use voxel for the static parts of the world and polygons for the dynamic parts, again using Final XIII as and example where most of the backgrounds you can only see but not interact with.

This has actually been done in games already. Outcast used voxels for the environment, trees, buildings and stuff and polys for the characters. The problem they had was that they had too use quite low res textures and stuff so the characters were not that detailed.

Still, I have to say that when I saw the game for the first time I was struck by the beauty of it, especially the softness and roundness and how organic it looked. If you can still find poly edges in the games today imagine how they looked like in 99...

 

[Platon]

I am cautiously optimistic, but the big thing is indeed how the system behaves with animated objects. To me it seems that the big breakthrough is mostly the system of finding the visible points not so much the point could/voxel tech as such.

Since they store the scene info in some format that is easy to search/retrieve visible/relevant points it would be interesting to see how it would react to objects that are not static. How do you easily file/store points for later use, when you don't have a fixed 3d space for them. A bit like his analogy of finding a book among a billion books where you have to read all the titles until you find the one you want, compared to a library with its filling system which is much easier to find what you look for. But how do you find a book that has been placed at the wrong place, then you have to go through all your billions books anyway.

Now, analogies never really tell how things exactly work, but still. I hope they do get funding and see if this tech actually comes to fruition, if nothing else so they can hire an artist or two, and make some more interesting demos with some animation...

 

[Laa-Yosh]

Honestly, there are just far too many unsolved problems and the benefits are superficial. Do you really want such a heavily instanced world? Even Carmack isn't pushing voxels now, they've probably realized that the drawbacks outweigh the benefits by far.

Polygon based data representations still have a lot more to offer, just take another look at Lionhead's Megameshes to see what's a far more likely new direction. This unlimited stuff is a tech we do not really need!

 

[archangelmorph]

Honestly, there are just far too many unsolved problems and the benefits are superficial. Do you really want such a heavily instanced world? Even Carmack isn't pushing voxels now, they've probably realized that the drawbacks outweigh the benefits by far.

Polygon based data representations still have a lot more to offer, just take another look at Lionhead's Megameshes to see what's a far more likely new direction. This unlimited stuff is a tech we do not really need!

I disagree somewhat.

After seeing some of the strides that have been made in compression & storage methods of sparse voxel octree data scene representations it's possible to see even more efficient representations of "geometry" in a virtual world than polygonal methods.

However...

I think the work has a *very* long way to go before it could ever be practical for real-world game usage & there are many many challenges ahead before we can really get there. Not the very least of which would be convincing MS, Kronos & all your GPU vendors that such technology is the future of computer graphics, sufficient for them to invest in designing their hardware towards it.

Also I think the progress these guys have made (if any as the overall tone of the videos they put out seriously undermines their credibility with me...) is very seriously overstated. Especially when throwing out statements such as "unlimited graphics power" which is bollocks.
"Unlimited geometric detail" (i.e. what the software+hardware does) & "unlimited graphics power" (i.e. what the hardware specifically is capable of) are clearly not the same thing & if they are stating as such, either they are full of horsesh** don't know anything about computer graphics hardware & real-time rendering (or at least the commentator doesn't) or they are deliberately making wild baseless claims in a bid to convince the uninformed that they've achieved something revolutionary.
Either that or it's some form of patent trolling...

 

[Laa-Yosh]

I think I've mentioned "drawbacks" and "benefits" and their relationship to each other...

Polygons aren't only about efficiency, but also about flexibility - which this technology can't yet offer, and voxels are the same with their own share of restrictions.

Again, look at Megameshes - efficient representation of a high level of unique geometric and surface detail, combined with streaming tech to work on a current generation console. Expanding that technology for the PS4/Xbox3 seems to be the logical next step and it will allow a significant jump in asset fidelity - while keeping all the flexibility we currently enjoy with polygons.
I still think it's a far more promising direction for research and development.

 

[Jedi2016]

The other thing they neglected to mention was that there are other things working to make our games look good besides polygons. Games like Crysis and BF3 don't look as good as they do because of poly count.. it's the shaders, lighting, and all the other little tricks.

Any CG artist can tell you the same thing. No matter how detailed and amazing the mesh, it'll look like crap if you don't know to make adequate use of it. And even motion pictures get away with surprisingly low-poly meshes sometimes.

 

[ultragpu]

I can definitely see this tech work well with certain games such as flight sim or ace combat, just imaging how uber detailed the terrain would look since they don't animate. But I hope they keep up the R&D for this tech and perhaps they would perfect it by the time we desperately require a new alternative to polygons.

 

[jlippo]

How would point clouds work with collision detection and actual gameplay mechanics?

Are there anything else the tech is good for than displaying single view?
If not this is not the way to go.

Voxels at least have advantages of ray-tracing and are easily streamed.

 

[whome0]

What the hell did I just watch?

The same as last year...

Sure you mean same as The Australian Game Developers Conference in 2003?

 

[dragonelite]

A interview didn't had time to watch it
http://www.hardocp.com/article/2011/08/10/euclideon_unlimited_detail_bruce_dell_interview

 

[Holborn]

What i don't understand is how they think this tech would perform in real world situations -like games.
Games are depended heavily on caching. You have a rock, you store it and then you place it multiple times on a map. With UD it seems that the same rock must be reconstructed/recalculated and re-applied every friggin' time. Do they use tables of "points" that they name "rock" or "tree", or what?
Also shading, lighting and skinning this would be a be*tch. Leaving no room for anything else.

They should go to Futuremark and request a Demo.

 

[Laa-Yosh]

Er, I think you have some misconceptions here.
They don't need to recalculate the same rock at all, they're using the same source data to display it a gazillion times. That's exactly the problem, that their entire game world is built like that, there's very little unique detail.
Oh and the other issue is that they can't arbitrarily rotate the rock, it'll always be aligned to the XYZ axes of the game world.

Shading it shouldn't be much of a problem, they could even render out a G-buffer and do deferred rendering. For now their source art has all the lighting and shading baked into it, but that's just a matter of increasing the art budget.

Now, skinning and all other kinds of deformations, that's the major problem and they haven't shown anything promising there.

 

[Laa-Yosh]

I really have to paste this because it's so unbelievable.

The problem you don't understand, is that point clouds don't need TEXTURES. Textures are purely a POLY thing. Point cloud data includes the texture as part of the point cloud, and also includes the COLOR.

You don't NEED terabytes of texture data. However, unlike a poly, which is basically three coords and a bunch of texture data, you instead have 10k point objects.

The key, AGAIN, is that he's not requiring the system to render the ENTIRE point cloud for an object, and then bi-sect, and tri-sect the result to get the 3D projection, he's culling the point cloud based on the pixel projection right from the start. That results, in the only data necessary to render the screen is only the points which are visible within the pixel frame.

A similar process for poly-casting would be to imagine each pixel as an individual 3d projection screen, and ray casting against the poly based upon what's in front of the pixel. Your projection is built into the process.

Carmack mentions these processes as REAR projection as opposed to FRONT projection. Most current poly techniques use a FRONT projection technique which requires a fully rendered item to then be projected from the object to the screen. Euclidon likely (we're speculating here based upon available information), is that he's using a REAR projection technique which limits the rendering to only those points of the point cloud falling into the search algorithms result.

Not saying REAR projection doesn't have it's problems. Shadows, smoke, and other environmental effects become more difficult since they are procedural systems which aren't part of the normal point cloud, and instead, modify the point cloud data. It's not an intractable problem, just one that requires more thought than current poly techniques do.

At the same time, as mentioned in the video, the physics problems become infinitely EASIER since your resolution is high enough to encapsulate some of your physics problems (bounding boxes, perimeters, and other poly techniques for finding edges). So, while making some things harder, it makes other things EASIER.

The biggest hurdle is going to be moving the mindset to a different series of techniques. Most of the graphics professionals in the medical industry are well versed in these systems since they are actually used quite effectively in systems such as your 3-D MRI scanners and LIDAR based rendering. We'll see some crossover here in those techniques are being applied here into the game-world...

 

[AlexV]

Yeah, let's not get into trans-site sniping now, mmkay? Also, on a general note, this is the internetz, expecting reason in the face of outlandish claims may be a bit too much for the general case. I think discussing SVOs is interesting, and quite B3Dish, so let's stick to that irrespective of what started this thread. Thank you!

 

[Shifty Geezer]

The problem you don't understand, is that point clouds don't need TEXTURES. Textures are purely a POLY thing. Point cloud data includes the texture as part of the point cloud, and also includes the COLOR.

You don't NEED terabytes of texture data. However, unlike a poly, which is basically three coords and a bunch of texture data, you instead have 10k point objects.

And they're ignoring completely that that the same terabytes of texture data is going to need to be included in the point data?

 

[Laa-Yosh]

Yeah, Bruce Dell has done some unlimited damage to the industry and the general opinions of the public. I hope he'll reap the rewards for it eventually.