Monday, February 4, 2008

Rendering complex procedural geometry.

Andreas Dietrich, Gerd Marmitt and Philipp Slusallek. (2006). Terrain Guided Multi-Level Instancing of Highly Complex Plant Populations. Proc. IEEE Interactive Ray Tracing. Related site.

RL Cook, J Halstead, M Planck and D Ryu. (2007). Stochastic simplification of aggregate detail. Proc. ACM SIGGRAPH, 119-128.


Unknown said...

Terrain & Plants:
This paper begins by claiming ray tracing can now allow photorealistic landscapes but their pictures are FAR from photorealistic. Anyway, on to the actual paper...

As I was reading this I actually was thinking about using Wang tiles to do part of their work and then *boom* I see that they actually did use them.

In all I don't have many comments on this paper because it never really struck me. What's their big idea? I see a couple of "new" ideas: lazy kd-trees and adaptive plant density. These, however, only seem slight adjustments on other ideas and the authors did not devote much time to them.

Stochastic Simplification:
Wow, I really have to show some respect to an academic paper that admits "Fortunately the exact formula is not critical." Props to Pixar.

The authors have an entire step dedicated to contrast preservation, yet entirely ignore contrast as part of the level-of-detail calculation. I wonder why they excluded it, they never really explained. Is it that, like blur, a useful model is hard to come by? Or is it better handled in conjunction with the rendering priority tweak or what?

Depth of field blur seems to be becoming a hot topic in graphics and it is discussed in this paper. I remember quite a while back that one of the Voodoo cards (I think the 4/5 series) had hardware blur as part of feature set. This was the only card I know of to include it. Granted this paper is ray tracing, but is the hardware model different from the software model for blur?

Stuart Heinrich said...

stochastic simplification - it seems that their automatic LOD method is novel, and does its job quite well. of course, it is also highly specialized, being applicable only to models that fit the property of containing a large aggregate "mass" of seemingly random triangles, such as foliage, hair, clouds, etc. It's interesting how they had to compensate for increased contrast, I wasn't expecting that to be a problem, but I see why it is.

Terrain guided instancing - I was a bit disappointed with this paper. I thought that they might also address the problem of realistic tree distributions. They basically just used kd-trees, extended to this purpose, which is pretty straight forward. For some reason all their tree renderings look very sparse, perhaps just a bad tree model. Maybe there could have been more variation between the trees, too. What bothered me the most though were the odd rendering artifacts occuring at regular intervals in image space. It appeared that they were not properly combining render buckets.