Comment by coffeemug | Hacker News Reader

Comment by coffeemug | Hacker News Reader

I think what they're doing is great, but I see two problems with their presentation. First, computer rendering techniques are extremely well understood and well researched. We've picked the low hanging fruit, much of the high hanging fruit, and everything in between. There is no "groundbreaking new technology" to be invented. They're converting polygons into voxels (although each voxel is probably a sphere for cheaper computation), and using software ray-tracing to render in real time. Since ray-tracing is trivially parallelizable, the multicore technology is just about there now. A 12-core machine will give just about 20FPS. The reason why they can get away with an incredible amount of detail is that ray-tracing diffuse objects is fairly independent of the number of visible polygons in the scene.

The second problem is that 10^4x improvement in level of detail does not mean 10^4x aesthetically pleasing (or in fact, more aesthetically pleasing at all). Ray tracing gets very expensive the moment you start adding multiple lights, specular materials, partially translucent materials, etc. It is very, very difficult to do that in real-time even with standard geometry, let alone with 10^4x more polygons. This is why their level doesn't look nearly as good as modern games despite higher polygon count (compare it to the unreal demo: http://www.youtube.com/watch?v=ttx959sUORY) They only use diffuse lighting and few lights. In terms of aesthetic appeal of a rendered image, lighting and textures are everything.

Furthermore, one of the biggest impacts on how aesthetically pleasing a rendered images looks is made by global illumination. That's also something that's extremely difficult to do in real time with raytracing, but is possible with gpu hardware with tricks. The trouble is, these tricks look much better than raw polygons.

Again, I love what they're doing. Real-time ray-tracing is without a doubt the future of graphics, but it would be nice if they were a little less sensational about the technology, and more open about the limitations and open issues.

This video is not that fresh. I have seen it already last year and the way how this guys speaks has not changed since then (speaks like a salesmen) ;)

Their technique is based on point cloud rendering. My supervisor has proposed already in 2004 how this can be done on a standard PC. Look at his PhD-Thesis [1].

This technique is usable for static objects as well as for dynamic objects; however, using it for dynamic scenes additional acceleration structures are required (besides of the Octrees) in order to dynamically change according to the deformation of the object. To clarify on several comments made here:

- this is a rasterization technique

- they used acceleration data-structures like Octrees, kd-trees, BVH, ... (which exactly they don't tell)

- the graphic "aesthetics" depends actually only on the artists and not on the technology itself, so not a good point

[1] M. Wand: Point-Based Multi-Resolution Rendering. PhD Thesis, Wilhelm Schickard Institute for Computer Science, Graphical-Interactive Systems (WSI/GRIS), University of Tübingen, 2004.

> I have seen it already last year and the way how this guys speaks has not changed since then (speaks like a salesmen) ;)

+1. I'd actually trust the company more if they had someone who spoke in measured tones and laid off the hyperbole.

But did your supervisor ever ship a product that did it?

Yes. We use his original software, XGRT, every day in our research group. All our research about point clouds and geometry is implemented as modules in this software. The software is open source and supports point clouds of really huge sizes, e.g. 80GB datasets, dynamically streamed from the disk.

And there are no moving objects in this demo. Voxel animations are much harder than poly animations.

Then there is memory. The elephant is looking great, no doubt about that. But I think you will need a lot of space for it. On a PC this might work, but I'm not sure it can be used on consoles.

That and I don't know of anyone (outside of using them as particles) that has successfully created collision models with voxels in real time. The collision calculations of just a few of their scanned rocks and a ground plane would be plenty complex - a whole scene perhaps even a little insane given today's specs.

They might have to resort to polygonal collision models in the same way that polygonal games end up using low-poly collision models (with the same pitfalls such as moonwalking, blocked projectiles or mystery-bouncing).

Yeah I think so. A tree-lookup can make things easier, but then you have to estimate the bounce direction.

But overall it's a great demo. And you see more and more games using voxels for surroundings. So ofcourse you could combine this technique with polys.

The more I think about it, the more I realize that of the three pure voxel problems (rendering, animation and collision), rendering is probably the low-hanging fruit - Though I'm not sure if creating a hybrid poly/voxel engine is any harder or easier.

Here's to at least getting a hyper-real Myst someday :)

Collision need not be at the same resolution as the rendering which should help. But animation is probably limited by how much of the scean changes. If it's a small percentage of the enviornment removing the old objects and adding the new ones should be reasonably fast. Or if it's a small number of objects you could always keep them in a seperate tree and translate a ray in the main scean with one for the moving object.

What about this: http://www.youtube.com/watch?v=Gshc8GMTa1Y&feature=relat...

As far as I can tell, they're not doing any voxel-voxel collisions. Mesh-voxel collisions sound much more tractable.

Sorry for the belated reply. Yeah, I saw Atommontage and it's quite interesting. Despite the realism of the track marks left and suspension movement, the truck still seems to exhibit some unnatural "floating" feeling. These are the very pitfalls of dealing with lower-res poly collision that I was speaking about. Kind of an uncanny barrier for movement.

I do feel that this author is a bit further along at something ship-able than the Euclideon folks are though.

They are using point clouds, not voxels. Animation should be easier without the grid restrictions of voxels, though it really depends on their acceleration structure (which they're not saying anything about, in addition to the specs of that machine they're running on which are undoubtedly beastly).

Agreed, not very impressed.

Realistic movement is a more important factor than detail, you can have 8 cubes move fluidly enough to immerse in. 1,0000,0000,0000 polygons seems un-important when you have pixel shader's and normal mapping.

Maybe they should of made duke nukem forever with this engine, as it seems like its from a similar era.

The new frostbite engine 2 a million times amazing, because of the dynamic lighting and radiosity. Plus the amazing fluidity of the animations.

voxel animation: http://www.youtube.com/watch?v=Tl6PE_n6zTk

This is somewhat the state-of-the-art in realtime ray-tracing (actually path-tracing): http://www.youtube.com/watch?v=qntMz2QPA_E using the Brigade renderer. It's using both the GPU and CPU.

That game looks like it has been filmed with a cheap digital camera with the ISO setting too high.

That's the neat thing about path-tracing, you can render in a very short time but get more noise. More powerful hardware or more time gives noise-free and unbiased image.

> They're converting polygons into voxels

Are you sure about that? He mentions atoms, but then the video also mentions that they're using procedurally generated graphics, which is something entirely different. Also, voxels would not work on the scale they are demonstrating, you'd need way too much memory.

I thought he said something like "you don't see polygon counts this high unless they are procedurally generated".

Actually they say they're doing 20-30fps on 1 core.

In terms of aesthetic appeal of a rendered image, lighting and textures are everything.

A thousand times this!

Reply on news.ycombinator.com