Any riggers here? ;)

[nodeway]

Hi.

Is anyone familiar with Smooth Skinning Decomposition or Deformation Learning Solver? Does it work? Do you think they produce effects good enough that you could use them in production?

I'm looking at next topic to implement in Houdini and Smooth Skinning Decomposition is one of the candidats (Continuous Visibility Feature being another one). First would be useful for is rigging, second for terrain generation tools and occlusion culling in game engines. I'm trying to decide on which I should spend my free time.

Have a nice weekend!

[Mathaeus]

About Smooth Skinning Decomposition from this article, I don't believe you'll find someone who is able to talk that much. I mean, entire thing looks like something still reserved for big facilities, with all appropriate NDAs and such.
More or less 'manual' methods of converting the arbitrary deformations into skinning, are well known for long time, let's say in old article about XSI and metal gear solid - basically it is a bunch of nulls, constrained to deformed mesh, used as skin carriers in next step, and that's it.
Of course, machine does the job x time better, but, there's another factor: (if I'm correct, in 3.2 of article) it seems they are proposing a new method of skinning interpolation, different than standard Linear Blend. If so, complete thing is able to work only where is implemented, Houdini in this case - but not in certain game engine. While big players are conservative, these days: for example, Dual Quaternion skin is removed from Unreal before few years (from what I heard, almost nobody used it), even DQ skin is an almost standard thing in 3d apps, today.
In short, thing is looking like something more in area of specialized solutions, like DQ skin, ILM Skin, some Disney variance with movable center of rotation, whatever....

It also depends on what is your goal, 1: to make easier life for some Indie user who's trying to animate in H, or to export something to game engine, or, 2: to get a job in big facility.

If goal is around 1, perhaps some additional automatic method of skinning wouldn't hurt, as an unavoidable, always not-so-easy task.

Very personally, I'll be more interesting in terrain generation, I'd believe it's better for H to be focused in areas where it is already (more) successful. For rigging, skinning and such, not easy to compete against Maya LT or Blender, these days.

[nodeway]

I'm asking because I saw this


and it looks cool.

Possibility to bake to bones automagicaly is really cool.

[Mathaeus]

I'm asking because I saw this

and it looks cool.

Possibility to bake to bones automagicaly is really cool.

Aha, yeah that one really looks cool, feasible, compatible with game engines, anything. There's only one thing that makes me worrying - for now I don't see additional, interpolating bones around knees or elbows - behavior of their outer sides could be a nice example of Linear Blend limitations - as name somehow says, in case of two influences, result is always somewhere on line between two moved reference (bind) poses, in case of three it's always inside the triangle, so on. It just can't go ''out'', so there is visible 'rounding' (or, collapsing) around knees, also it should be much more noticeable on high angles. In other word, if it's possible for this solver to work with any combinations of bones (or at least with few ''interpolators'') - that would be great.

P.S. facial rig seems to works much better, I'd believe because of chance for more uniform distribution of bones, also, maximum delta angles are relatively small (that's my personal theory...)

[rray]

My guess is if it's a typical machine learning system with a set of inputs and desired outputs, and the parameters you tell it to optimize are the main bone weights, it would probably mean you get the weights for helper bones like roll bones for free.
If their angles are not part of the input set their weights should just get optimized along. At least I hope so too.

Really interesting technology, and since "will it work in game engines?" is becoming more and more important, I'd probably pick that of the 3

[nodeway]

My guess is if it's a typical machine learning system...

On topic of machine learning, is it used in production?
Is something like this:

or this:

or this:

or this:

or this:

is used in any way in production?

What are the most crazy rigs you have seen used in production?