Unlimited Detail, octree traversals

I just have to comment on this one

10k colored points per object... That equals texture detail comparable to 100x100 texture unwrapped around the object (100*100=10k). Assuming the object is somewhat box shaped, we can approximately split this detail to it's six faces. So one box side has 1666 points. If the points are equally spread around the box face you get (color) detail comparable to a 40*40 texture per box face (40*40=1600)...

Assuming I need to have the (essential) toxic waste containers in my game, with signs and the warning text at each six box sides, the 40x40 points (color values) per box side is nowhere enough. Current games use 512x512 textures to make their toxic waste containers look good (sharp surface texture detail). That's 163x more color data than the 40x40 point cloud detail can offer.

So you either need to bump up the point amount to match the texture detail of current generation games, or you need to have more (separate) color (texture) data than position (point cloud) data. Separate color (texture) detail and geometry detail has it's advantages.

I have done some point cloud / voxel rendering research myself (I coder has to have some hobbies after all). And personally I feel that storing texture coordinate data (a big virtual texture) to the points/voxels instead of color/normal/material data has it's advantages. Instead of doing some screen space gap filling / parameter smoothing (to smooth the "point sampled" point/voxel material parameters and normals), you can just interpolate the texture coordinates. The virtual texture contains high quality color, material properties and normals that can be filtered properly. And the artists can freely create higher (or lower) texture detail compared to geometry detail, and curved surfaces have perfect specular highlights (virtual texture stores object/world space normals). This kind of point cloud rendering could also combine really well with a g-buffer that only stores texture coordinates (to a virtual texture cache). The voxel renderer would only need to output texture coordinates and depth values to the g-buffer. All other parts of the rendering pipeline could be done efficiently with current rendering techniques.

 

But the question of animation / geometry deformation would still remain unresolved, am I right?

I'd still like to see Lionhead's stuff developed further where it can build ingame meshes that can be tesselated and it can generate displacement maps in addition to normal maps. Current consoles can't possibly run it but next gen hardware should be enough to make the rendering part work. It's the content creation aspect that could provide a serious competitive advantage.

 

Nah, you don't understand. This is UNLIMITED detail. Just increase the number of points in the barrel to 32 billion per face and the colour fidelity will be extraordinary, without needing to store 32 billion colour values per face because....um...because it's UNLIMITED DETAIL!

Perhaps the most worrying aspect here is a total lack of the most basic understanding on the part of believers. That they don't think to themselves about the requirement for information and how that is to be stored and accessed boggles my mind, and just goes to show how all too often people take things on faith without any of the most essential questions being asked before hand. The most rudimetary observation of this tech is, if you have highly detailed point clouds, you need storage for that data. Now if they were talking about HOS rendered as point clouds or something, they might get somewhere, at least regards unlimited 'polygons' in games. But unlimited information is...well, I suppose it's a great idea for getting dumb investors on board!

 

It worked with the australian government too, Euclidean has managed to get 2 million dollars from them.

 

Governments are always completely clueless when it comes to technology. Ours wasted countless billions on a national health database without getting a single usefull product in return, where a decent collection of university undergrads could have developed a suitable system. Perhaps one of the most lucrative scam targets is governments because of this?

Before anyone invests, don't they ask an independent consultant on what issues this tech faces to make a wise investment?

 

Not sure here in holland the goverment had invested 5 million in a national police system.
5 top IT companies said this system the goverment was investing in would never work but they still did it. And now if im not mistaken they are going to look at the system to replace it.

That is why i think such contracts has to be friends giving contracts to friends.

 

The problem with governments at least from what I saw in my country is that they usually go with projects where some public employees, usually in the higher positions will gain from it. They could either get money behind closed doors, get money from the cost incurred or the organization/person/group that executes the projects have some connections with some public employees in key positions.
In other cases they take the offers with the lowest possible cost instead the one with the most beneficial results.
Some people in the government dont care about the long term benefits for society. They look for good excuses to take up projects that will of course generate money for some pockets.

 

Sorry. I will keep things separate from now on.

 

Yeah. But the reason for that is that most technology is transparent. It is designed to be invisible to the end user, so expecting them to understand what happens underneath is counter-intuitive...

 

but I can't wait for the next wave of "procedural content will take over", a short while after Rage is out I guess.
then we may be given a new look at the tech : UNLIMITED PROCEDURAL DETAIL!

at least it's not a fractal viewer written in visual basic.

 

Unlimited details for free sounds as believable as infinite growth within a finite system. Shifty you should be less tough with the internet, we are all "trained" to accept some facts as true

what really surprise me is the founding if it's a complete scam. I mean it doesn't take a PhD to check the size of the data used by their demo. Governments can be scary.

 

My memory is not what it used to be, but didn't they (under a different name) get another grant some years ago?

Still, it can't be as bad as the "Horvath Hydrogen Car" debacle.

 

I hope so to, the amount of abuse that people are getting on other sites for pointing out the flaws with "unlimited detail" is absurd!

 

ignoring the marketing term 'unlimited'
I just want to point out that DNA seems to store a lot of info. or say a curve you can either store it as a math equation or as 20trillion polygons, the equation uses far less storage & has potentually better results.

not saying I believe these claims from this company though (havent really looked at them)

also heres some work of mine from ~8years ago

ok vegitation/rocks are simple
but here everything is generated anew per frame (worsecase senerio), but good since everything can be dynamic, contrast this with our nearly static game worlds of today. everything is generated down to a level of how close the object was to the camera, thus the same tree can have 10million polygons or just say 100

 

I already mentioned that:

There's been no talk of representating the data that I've heard that suggests anything beyond a mass of point data, and evidenced by repeating the same data over and over again in the demos. There is a possible non-polygon engine that uses HOS and breaks them down into point clouds, but HOS isn't progressing very rapidly, and we have enough polygon power to render them as triangle now anyway. A 'perfect' triangle engine that renders one triangle per pixel would only need 2 million tris a frame, 120 million a second at 60hz, when hardware is capable of billions of triangles! We just have to use them other ways too.

I don't see triangles as a dead end though. HOS and displacement maps are the future for high detail (and procedural content modification/generation ).

 

http://en.wikipedia.org/wiki/NV1

 

Never heard of quadratic texture mapping.

 

In one of the presentations at SIGGRAPH 2011 (or possibly HPG - it's all a blur) it was stated that this nearly killed NVIDIA.

 

As various articles on the web suggested when I started Googling, it was because DX took a polygon approach and NV1 wasn't a DX accelerator, hence devs avoided it. I'm curious what exactly the tech was (some sources suggest it was NURBS) and how well it performed. This whole topic of super high detail, beyond just the recent Euclidian noise, could do with some fleshed-out back-story on the choices made regards polygon rendering and asset popelines and how that has or hasn't helped the adoption of HOS. As many are saying, Euclidian aren't trying anything not tried before. The fact Bruce comes at this ignorant means he doesn't know what has been tried before.

 

Watched it. It's useless! I like the guy, and how he made his way out of a dead-end existence. His alternative view of graphics has merit and he deserved a break. He doesn't answer any of the vital questions though, and he shows a fair bit of ignorance that's gone from quaint naivety to being somewhat embarassing. LOD = level of distance? And the examples Notch gave weren't explicit but about voxel data, yet Bruce just says, "we're not doing that (raytracing etc.)," without recognising the complaint. The subject of storage is raised and Bruce answers saying they have no problem with storage, but then they are using only a handful of objects, all regularly placed. There's nothing to indicate his sorting and search method will scale up suitably quickly. There's nothing mentioned about decent animation when a few basic wobbly objects are shown. There's nothing said about shaders and lighting except "that's coming". I have serious doubts he'll be able to pull it off. Basically the tech is a nice, rudimentary graphics demo, like the SAARCor detail demos, not bothering with any of the essentials required for a game. We need facial animation, dynamic lighting, physics interactions, and huge variety of content, none of which are a good match for compressed databases of points. Without showcasing any of that, or saying anything more than, "yeah, it's coming, trust us," I can't see any reason to believe in this tech. It's a spacial sorting algorithm on compresssed data, he's confirmed, fetching one datapoint per pixel (how's he going to deal with AA other than post or supersampling?), which isn't new, yet with his head in the sands Bruce believes it is.