Proceedings of the 16th annual conference on Computer graphics and interactive techniques
Pages: 41 - 50
Year of Publication: 1989
ISBN:0-89791-312-4
(ACM Digital Library)
http://doi.acm.org.www.lib.ncsu.edu:2048/10.1145/74333.74337
(ACM Digital Library)
It was interesting to see how modelling terrain was done back when there was relatively little work done in the field. The authors of this paper had to use their knowledge of mathematics and physics to create a new model of synthesizing terrain. The paper really shows its age when the authors talk about 2 MB of memory being a lot. That being said, the pictures shown in the appendix look fantastic for 1989.
ReplyDeleteI thought that the paper was not well written. It relied too much on reciting formulas and didn't have any sample pictures throughout. This made it hard to understand. I would have liked to have seen more sample pictures showing how a terrain would look with current methods, and then how it looked after they applied their new model.
The authors said that a future improvement would be to divide sediment into the different elements that compose it, and then assign each element a value that denotes how hard it is. This way one could get a more accurate model of how the water that carries the sediment erodes rock. Also, the grid tracing algorithm that was used in ray tracing could be replaced with the kd-tree.
I imagine that this model is outdated today and doesn't have much use, but it was probably used widely for generating terrain in artwork and videos during the early nineties.
The authors presented an interestingly physical approach to modeling erosion, that is, actually imagining that water is being poured onto the surface of the terrain and calculating erosion and deposition in frames. In the world of "If you can't make it, fake it," I was surprised at their high "make-to-fake" ratio for erosion. Also, the technique for raytracing a height map was new to me, and it make a lot of sense. On the other hand, most of the modeling section went over my head because I don't have a strong background in fractal math. In the end, though, I thought their results looked much better than the few pictures of fBm surfaces that I found online.
ReplyDeleteI thought the paper was well-organized and the authors explored the topic in a logical fashion, starting with modeling, moving onto refinement with erosion and thermal weathering, and finishing with rendering. However, they could have saved me a lot of trouble on Wikipedia by defining their terms on first usage instead of waiting until the middle of the paper to define, say, lacunarity.
Since the authors were trying to simulate real terrain better than existing methods, I think their conclusion would have been stronger if they had included a comparison of their results with existing results as well as real terrain to demonstrate the strengths of their method over others. One thing they didn't cover that I was curious about regarding their erosion algorithm was the equivalence between passes and geologic age. For example, if I wanted to simulate a hundred thousand years of erosion, roughly how many passes should I run? Perhaps this conversion is not as simple as it seems, though.
Obviously, fractal generation, erosion and raytracing of terrain has applications in the world of entertainment (movies and games). Perhaps when the algorithms grow more sophisticated (if they haven't already) they could be used to model and predict actual erosion for geologic study. As for the raytracing algorithm, I would be interested to see whether or not recent general-purpose raytracing methods (such as SAH kd tree) obsolete their specialized technique. If not, their technique could be used in conjunction with recent general-purpose methods for complex scenes involving a height map as well as other objects.
"erosion#3" is an implementation based on this paper,
ReplyDeletehttp://www.bealecorner.org/best/landscapes/index.html
Here is a more recent work on simulated hydraulic erosion:
ReplyDeletehttp://evasion.imag.fr/Publications/2007/MDH07/
The authors proposed what I think was a really novel technique by the time the paper was published. They came up with a method that gave the terrain synthesis a different approach since it addresses the symmetry of the fractal terrains surfaces and the lack of erosion features as their main problem. It is really interestung how they formulated an erosive process that is divided in two phases, the hydraulic and non-hydraulic categories.
ReplyDeleteI consider that the paper was hardo to read since they introduced new terminology along the text without previously define it. The authors have written the document emphasizinf the mathematical and physical principles that support their methods, which made me feel lost in the reading sometimes. However the purpose of each step of their algorithm was pretty straightforward and despite all the mathematical approach the steps of the algorithm were fairly easy to understand. They have a nice preamble where they defined and gave detailed information of the related works that had been done in the same reasearch field by the time; this helped to understand that the author's method was addressing specific topics that had never been solved very efficiently.
They presented a grid based ray tracing technique, which nowadays is not the most efficient approach since it implies the calculation od the intersection point between the interval of altitude and the ray. Their method can be improved by implementing SAH or KD-Tree.
Lastly, terrain synthesis obviously has a great field of application in films and video games. It can also be extended to engineering or any other field that requires topographic representations of terrains.
The paper describes a very interesting approach to terrain generation. The example plates they include show a clear improvement over the fractal patterns they are based on. I have no doubt that these methods have been imporoved upon in the 20 years since this was published, especially with the huge advances in hardware and rendering methods since then. This does make the sections focused on implementations a bit outdated, since they detailed limitations which really don't exist anymore (at least on that scale).
ReplyDeleteThe paper itself was clearly written from a more mathematical perspective. I'm not particularly familiar with fractal math, but I do have some background in signal processing. The authors used signal processing terms and methods rather heavily (especially in generating the base terrain). For example, they focused on the ideas of power spectral analysis and stationarity (among others) to explain the noise generation methods. At our university at least, these methods are much more common in electrical and computer engineering. I'm not sure if this is a change in focus over time or just a difference in programs (or both).
I thought the organization of the text was good. It followed a logical progression through each step of the process, and I thought all the relevant concepts were explained pretty thoroughly (even if they were sometimes in purely mathematical terms). One thing that was rather annoying was having to reference the plates located at the end of the paper.
As was stated with the other paper, terrain generation has a very wide range of applications, from games and animation to simulation. I think these methods could probably be improved by adding a focus on parallelization for modern graphics hardware, using more current rendering techniques, and using the additional headroom of current technology to make more detailed analysis (like their example of varying hardness of sediment).
I really enjoyed this quote:
ReplyDelete"our ongoing quest for more beautiful and realistic forgeries of nature"
The reliance on height fields makes it impossible to generate any landscape that can't be represented as a height field. Stone formations frequently have negative space below the top point in a height map (an overhang, essentially). Water routinely carves through rock and dirt, creating obscured or even underground streams. Typically these are limited to small scales, because the overhangs collapse if they grow too large, but there are some large and striking counter-examples:
http://gallery.hd.org/_exhibits/places-and-sights/_more2000/_more06/Australia-Western-Australia-Hyden-the-Wave-Rock-15m-high-overhang-on-side-of-Hyden-Rock-vertical-rusty-red-ochre-and-sandy-grey-streaks-of-carbonates-and-iron-hydroxide-about-340km-east-of-Perth-1-RAF.jpg
http://science.nationalgeographic.com/staticfiles/NGS/Shared/StaticFiles/Science/Images/Content/arizona-rock-705695-sw.jpg
These phenomena would require some additional representation beyond a simple height map.
I would be interested to learn if anyone has tried applying reinforcement learning to this problem. Train it with some real elevation data, feed it treats when it gets close to the real thing.
The hardest part of the paper to understand were simply the unfamiliar algorithms. I very much aprpeciated the authors' summaries of the most important algorithms (fBm, Poisson filtering, midpoint displacement, etc).
Plate 9 is pretty impressive. The flat pool of sand or water is very convincing.
--
mwc
The idea of simulating hydraulic and thermal erosion
ReplyDeleteboth interested me and struck me as a novel idea.
The ability to depict the effects of nature over time will greatly
enhance the authenicity of rendered landscape images.
The authors presented their points in an organized fashion. At
times, I found myself getting lost, but overall, the authors did
a great job in telling us the topics they would be discussing and
then explaining the topics. The authors also tied their
different ideas together well. I was able to follow, for example,
how noise control tied into simulating more natural
terrain.
The only area of their method that might need
improvement is the use of a memory bound algorithm.
With Grid Tracing, if you are dealing with
a large amount of data, the time of rendering will
cost a great deal.
The ideas presented in this article could help scientist
study nature better and could help them analyze
the world we live in more efficiently.
While the paper is rife with terms and concepts that are novel to me, I must confess that I don't find them even remotely interesting. Many of the terms and phrases that are used throughout the paper were simple enough to understand once I bothered to look them up, but mostly they worked as a barrier to full understanding of the paper. But this is clearly no fault of the authors. The paper does seem both well-conceived and well-written. Perhaps it would have been more of an engaging read if they included a few more childishly thrilling bits like this: "... as B. B. Mandelbrot has pointed out, the fractal dimension of the Himalayas is approximately the same as that of the runway at the JFK airport".
ReplyDeleteJCS
I find it amazing that they were capable of producing terrian this realist back in 1989. I assume they were only rendering it in static images, but it's still very impressive. I would believe their concept to very novel except they mentioned that another team elsewhere came up with basically the same concept while they were doing this so it must have been the logical next step.
ReplyDeleteLike the paper before I would need to go read vast amounts of other work to be a ble to implement their concept, but overall I understood this fair well considering no experience in the field.
I felt the paper was well written. They started you out with a glossary of terms they would use that you might not know. The laid out what they were going to specifically talk about, and then did so in easy to distinguish sections.
Clearly their methods can and probably have been improved - the idea that something cannot be improved would lead us to stop trying new things.
This method would have been applied to terrian that you needed to get closer to as the old methods provided a good enough silouette type model, but this generates much better detail.