Hm, you mention that the scalar field is really easy to modify.
What kind of things do you think are possible for modifying this scalar field? Do you think it’s possible for us to define that we, for example, only want really simple face-loops around the arms and the legs? Or perhaps that one part of the generated mesh has higher mesh density? Or any of the other needs that have been mentioned in the initial PDF in this thread?
Perhaps, if we think this out carefully, the work you’ll do this GSoC can become a great base to create very good retopology tools upon!
Predictability is nice, but the more important factor is how controllable it is, and these are not quite the same thing - the tool is not useful if the resulting topology is predictably bad. Not that I’m saying your implementation will necessarily be uncontrollable.
Most of my skepticism stems from the fact that the examples in the paper are really poor - none of them are hard surface models, which immediately rings alarm bells. All of the source models are low poly (relatively speaking), unlike the multimillion poly models we’ll want to use this on. All of them consist of long tubes and don’t show the joints clearly, almost as if the author was hiding them. All the results have ngons or triangles - cleaning up the mess manually later would probably take almost as much time as full manual retopo.
Now, a little ngons or some bad topology here and there wouldn’t be so bad if this tool is just meant as a means of getting a workable basemesh for multires. But with Blender being an animation suite and not just a sculpting program, it’s important that the remesher is capable of producing animation-quality topology. Which basically implies control over every pole and loop. This is why I’m a fan of the semi-automatic patch-based remeshing and not terribly keen on the zimmer-bommes papers. And the paper you chose is just a mystery to me
whichever paper is used, i will continue to have interest in the development of this project.
Maybe not as a final retopology tool. More like a remeshing tool in a sculpting pipeline. This is perticularily important right now, as blender’s dyntopo is extremely slow once you hit the half million trianges mark. It gets completely unusable.
The current remesh modifier is not very good at keeping the shape of the object.
For that reason I am still using 3d coat or zbrush for sculpting. Blender is just not good enough to handle dense geometry yet.
At least not on my 4gb duo core intel laptop with an ati radeon 4650 card
Having checked both posted papers I have to confirm that implementing those is no small matter.
I used to joke that I would request an honorary doctor’s degree if I ever implemented any of those :P.
For GSOC the selected paper is ideal in terms of scope. Also noone really prohibits someone to implement another algorithm later.
@blurymind, that paper is just an optimization/extension to one of the papers posted initially.
whichever paper is used, i will continue to have interest in the development of this project.
Maybe not as a final retopology tool. More like a remeshing tool in a sculpting pipeline. This is perticularily important right now, as blender’s dyntopo is extremely slow once you hit the half million trianges mark. It gets completely unusable.
Blender dyntopo has been optimized on recent builds. 1 - 1.5 M faces are possible now, maintaining smooth sculpting. Of course you need a little better machine than a core2duo 4 GB RAM
In a sculpting multires pipeline however, you won’t be happy either. Such autoretopo meshes won’t work well with the existing multires modifier. You will face a dead end, unpredictably.
On a core2duo, you shouldn’t expect much from blender sculpting,
You mentioned 3dcoat, I doubt if it also performs well on your core2duo 4 GB ram. It does not.
Zbrush is your one and only real option.
I won’t say IMO because it is a fact, well known around.
My desktop machine was in the service for a couple of weeks
I had to work on my old laptop (core2duo, 2GB RAM, ATI)
ZBrush was my only option.
Well, i was just wondered there is a function in blender “tris to quads”, that one could use for model improvement.
For simple just take a human model or so, and then it removes a lot of tris; but not all while some of those are so obviously simple to solve.
Like going from vertical 3 squares to 2 squares, like i pictured below that shouldnt be too hard to solve i believe:
Note However i also think such tool should keep respect to the symetry of models, especial humans etc.
so it should kind start searching for triangles from the aoutside to the center or so. Nothing that resolving tries around the center is more easy then in other parts.
For the programmer =>
Basicly i think here one should take the triangle points, and count on each point how much edges it crosses, based on that the yellow made figure can easily be orientated.
And while going from 3 to 2 squares is easy but there must be more of these examples, take a pen and paper and a model; look for tries that wont get resolved by that function, see what goes wrong and then try to fit in an edge corner count to resolve it with a known solution.
Did everyone forget this paper was already part of GSoC 2010? It seems like there was never any actual work done on it, and the wiki page is now empty…
