Johannes Kopf Chi-Wing Fu Daniel Cohen-Or Oliver Deussen Dani Lischinski and Tien-Tsin Wong. (2007). Solid Texture Synthesis from 2D Exemplars. Proc. ACM SIGGRAPH. Related site.
Heeger DJ and Bergen JR. (1995). Pyramid Based Texture Analysis/Synthesis. Proc. ACM SIGGRAPH, 229-238. Related site.
2 comments:
Solid Textures:
Local and global characteristics are the two important things one wants to capture in synthesizing texture. Many methods handle only one, or one and a bit of the other, but only recently have good techniques been developed to handle both. This paper presented two ways to match global statistics (histogram matching and PCA reduction) and one for local (L2 neighborhood comparisons). Histogram matching is hardly used in the tex synth literature I've read and it is quite confusing why. It seems like a very easy way to capture some global information; not hard to implement and isn't computation intensive. These authors claim it gives better results, so why haven't others widely adopted it?
The main result of this paper is obviously 3D textures. One thing I noticed was while I was looking at the Chinese dragon in figure 5. I am a little confused as to how the synthesis uses the feature map. The authors say it provides non-local information, but what does this information correspond to and how does the algorithm use it?
Heeger and Bergen:
This paper's main contribution is the use of pyramids in synthesis. As mentioned above, capturing global characteristics is very important to producing realistic results. The pyramids presented here are one of the first attempts at capturing this information. In reading other literature, I've seen the Laplacian part of the pyramids excluded, leaving just the low-pass to be used in the synthesis. This is apparently because the Laplacian is expensive and doesn't contribute all that much to the results. It makes sense on a very high level to me because the low-pass contains course information that the pyramid eventually refines. What exactly would the Laplacian add to this?
The histogram matching method was pretty interesting. They tried to make it "physically based" which I think is mostly hogwash, because the use of gabor filters and brute force testing doesn't really have any fundamental difference from conventional image processing techniques optimized for computers. However it is interesting that the resulting patterns look sort of like natural animal camouflage. Also I can see that the gabor filters have prevented them from being able to capture realistic edges. That, and the fact that it is very histogram based. Their method seems very similar in principle to the monte carlo method used by Wei, except that Wei's method can more faithfully capture structure. Therefore, I would say Wei's method is superior, but this one has the advantage of being a lot faster.
The solid texture synthesis paper is basically the same algorithm is Wei's extended into 3D, and with some fancier math. The results are astonishing, certainly one of the most impressive paper I have seen on texture synthesis. By being iterative, it has significant advantages over all the procedural methods for 3d texture generation that are really limited by the need to find equations that mimic the texture. This doesn't. It can just do anything automatically, and from exemplars, so I find it very very cool.
Post a Comment