Slashdot Mirror
Mesh Compression for 3D Graphics
IanDanforth writes "A new algorithm that uses successive approximations of detailed models to get significant compression has been revealed by researchers at The University of Southern California. Just as MP3s remove high frequencies we can't hear, this algorithm removes the extra triangles in flat or near flat surfaces that we can't see. Experts in the field are giving this work high praise and imply that is will be immediately applicable to 3D modeling in games, movies, CAD and more."
I think this is interesting, but the analogy drawn between MP3s and this 3d-object compression is a bit strained.
The MP3 compression routine revolves around 'frequency masking' much more than it does "remov[ing] high frequencies we can't hear". Most of the work in MP3 is done through 'frequency masking'. That is, imagine a graph of the frequencies being played at any given time- find the high points, then draw sloping lines down to either side of those points. Humans can't hear anything under those lines- they're 'masked' by the nearby strong frequency.
Nothing very much like that goes on in this algorithm. There might be some other mesh-compression-analogous process that goes on in MP3 that's like this, but that ain't it.
Sorry to nitpick, but I figured it's important that
1. MP3 compression is not just simply throwing out high frequencies (a lot of these are actually retained) and
2. This isn't anything analogous to that, anyway.
Looking over my post, I'd have been fine if the submitter had said "Just as MP3s remove frequencies we can't hear, this algorithm removes..." but that's not very descriptive anyway.
RD
The actual paper can be dowloaded from here.
-jim
This is a great way to minimize scan data, but it isn't as useful as the article makes it out to be. Most modeled 3d objects are as low resolution as possible. Shrek has as many polygons as he needs to have, to take away some, or swap their location would destroy the model. For instance, I am a Modeler/TD and most animable character models have 5 divisions, or 'loops' around a deformable joint. Any less would not allow for the deformation control we need. As with most background scenery, it is modeled by hand and as low resolution as possible.
This could come into more handy later if it is built into a renderer.
A subpixel displacement renderer that can nullify coplanar polys in this way (though there arent that many usually in detailed oranic objects) it could speed things up quite a bit.
The article is short on technical details but...
h tm is in wide use. In fact, pretty much all 3d gemoetric level of detail techniques rely on collapsing "flat" areas. The source data for the geometry can also compress geometric data with stuff like NURBS and other parametric surfaces which is probably much better than some sort of lossy compression. With the coming "DirectX Next", OGL 2, and newer video cards, parametric surfaces (read: infinite curve detail) will easily become the norm.
While the algo may be new, the idea certainly isn't. Direct3D has built in support for optimized meshes, the ROAM algo http://gamasutra.com/features/20000403/turner_01.
http://brandonbloom.name
If you actually read it, it would be pretty obvious why this is new...sheesh!
Also, game data is built of far fewer triangles and in a much easier form than raw data read from a real-life source. (such as a laser range finder)LOD mesh reduction is usually done by full or partial MANUAL selection.
This isn't about compressing the data required to store a mesh, although it will help.
This is about reducing the complexity of meshes so that they can render faster.
autopr0n is like, down and stuff.
While I can't say for sure that nobody has used this method before for 3D models, the article seems to suggest that this is slightly different than using differently optimize models. Instead, this seems to be a way to optimize the models so that they look good up close as well.
:-)
The concept of lossy compression of 3D models might not be new, but that doesn't mean that the method for doing it isn't.
Also, even if the problem were trivial for 2 dimensions, it wouldn't neccesarily be so in 3. The 2 body problem has a simple solution, the 3 body problem has no solution in elementary functions. Random walks are recurrent in 1 and 2 dimensions but transient in 3 or more. I can think of several other mathematical examples where the difference between 2 and 3 dimensions (or 2 and 3 objects) changes things completely.
Don't judge unless you know you understand the subtleties of this algorithm compared to others
Un-disclaimer: I'm currently pursuing a PhD in machine learning.
Yes, it is new. First of all, y'all need to read the article and find out how.
It is for two reasons, both of which are stated:
The Desbrun team's novel approach comes from the seemingly unrelated field of machine learning...
Machine learning: getting a computer to generalize (invent hypotheses) given data instances. Work in machine learning has proven that generalization and compression are equivalent. That someone has applied those ideas to 3D model compression is at least notable.
We believe this approach to geometry approximation offers both solid foundations and unprecedented results...
In other words, it's not using some hacked-up heuristics. The bias behind the generalizations it makes are solidly described, and can be tuned. Machine learning consistently beats heuristics in accuracy, so their expectation of "unprecedented results" has a good foundation.
I got my Linux laptop at System76.
Read the fine article. You are correct that mesh optimization has been a most popular MA/PhD thesis subjects for over two decades. Which is exactly why someone comming up with a method that is an order of magnitude better than any other previous method is big news.
Also for all those questioning it's usefullness, you need not look any further than 3D scanning. When it comes to detailed models, very few things are done by scratch, instead the are digitized using one of many scanning techniques. This model is then massaged by hand by an artist. This technique would allow you to get a much better first cut, saving time for the artists.
Lastly, quake and others generated meshes from smooth NURBS objects. This is quite different, and much easier than generating one mesh object from another. Those tequniques are not usefull for scanned objects where you start with a dense mesh object.
Skimming the article, this just seems to be polygon aggregation on the model ( not HSR, which is certainly not what grandparent was implying anyway ). It's certainly not a method for compressing the stored mesh, it's just discarding arguably redundant detail.
Desbrun explains that his accomplishment was to simplify such a mesh, by combining as many of the little triangles as possible into larger elements without compromising the actual shape. Nearly flat regions are efficiently represented by one large, flat mesh element while curved regions require more mesh elements.
( My emphasis ). I was pretty sure this was nothing new, although I'm sure a general case algorithm, let alone a fast and accurate general case would be novel. But I was writing polygon aggregation code for my undergraduate computer graphics subjects ( much simpler meshes though ), and I would expect anyone with any CSG education to not confuse the subject matter with an actual storage optimisation.
One god, one market, one truth, one consumer.
Wow. That's pretty far from what "NP hard" actually means.