Technical Animation S21

Final

2021-05-14T00:00:00+00:00

Final Project Presentations

Our Presentation

Slides

Lecture 26

2021-05-05T00:00:00+00:00

Learning

Thoughts

Covered a variety of learning applications in simulation & animation
I still think the most compelling uses are seen in character animation, as mo-cap data sets provide rich basic data that can be intelligently interpolated & mixed by learning algorithms.
Still, learning for up-scaling simulations makes a lot of sense. Currently we’re seeing the wide adoption of deep up-sampling for images (NVIDIA DLSS), which is really just up-scaling a light transport simulation. So, it makes total sense that we could also apply these techniques to fluid/smoke or other physics simulations. I wonder if it is viable to train an entire physics and rendering system end-to-end, now that differentiable path tracing is also reasonable.

Lecture 25

2021-05-03T00:00:00+00:00

More Character Simulation

Thoughts

We only talked about “Fast and Flexible Multilegged Locomotion Using Learned Centroidal Dynamics”
The centroidal dynamics part basically meant modelling the character (at a very coarse level) as a simple inverted pendulum. Using this relatively simple model as a baseline that their controller can attempt to follow under simulation (while the details were quite complex), was a pretty elegant idea and lead to good results.
Their treatment of jumping and walking on un-even terrain was handled well, as the leg movement was adjusted appropriately to maintain certain speeds while taking physically realistic gaits.
I think I need to read the paper to really understand the details of their motion planner, though.

Lecture 24

2021-04-28T00:00:00+00:00

Character Simulation

Thoughts

A lot of work on physically simulated character animation is a bit less…physical…than I would have expected, as many papers focus on matching physical behavior to stylized or motion captured data. I guess the simulation is more of an extra policy to assure physically plausible movement (no intersections, anti-gravity, balancing) on top of art directed movement.
I do still quite like the classic analytic legged locomotion animation - something about it just seems cute and springy, albeit mechanical and unwavering.
The more generalized control models for e.g. biped walking involved a lot of classical control theory, which makes sense at a high level but I can’t say I understood all the math.
The various papers on learned controllers I found most compelling, as they tended to give very organic behaviors that are physically plausible due to the simulation. I really can’t wait until the learning based methods (e.g. phased functioned NNs) make it into interactive applications and games.

Lecture 23

2021-04-26T00:00:00+00:00

The Human Body

Thoughts

Hands are incredibly complicated - it makes sense that nothing less than the full muscle and bone models give really good results. The full body models that are based more on shape matching & mocap data were also pretty impressive, although they tended to lack the detailed dynamics of muscle movement, and seemed incompatible with joint capture of (loose) clothing.
It’s interesting that a lot of this research still builds on the basic building blocks of linear blend (or dual quat) skinning, blend shapes, and rigging, although the data sources range from mocap to MRIs.
I assume most of the soft tissue & muscle models require too much computation for real time animation, but I wonder if these will catch on in film - but also perhaps they are too close to the uncanny valley when paired with typically less compelling simulations like facial structure.
The paper on jointly learning clothing behavior as well as character pose might the future for animating clothes without cloth simulation. Their extension of skinning weights to 3D space via their neural net was a pretty cool idea.

Lecture 22

2021-04-21T00:00:00+00:00

Faces

Thoughts

I saw Yaser Sheikh’s keynote presentation at HPG 2020, which covered most of Facebook’s recent work on metric telepresence. Their encoder architecture was pretty understandable, and surprisingly straightforward going from on-headset video feed to the encoded representation, and then decoding to model deformations and view-dependent texturing. I wonder how extensible it is, though, since it seemed to need training per-subject. It is quite cool how well we can infer the whole face geometry even when occluded by a VR headset.
The facebook reality group also worked on eye texturing and tracking, which seemed like an even harder problem - they do end up using a separate model for eyes specifically, since they must be independently tracked by the headset and need very detailed texture. The learning architecture was pretty similar to the original work, but introduced extra loss for gaze direction and decoded both face and eye meshes/textures.
Lastly, the neural voice paper gave reasonable results from pure audio, which was pretty impressive. However, the animation was still a bit lackluster - certainly not yet lip-readable, and doesn’t capture broad facial movement. I imagine it is very convenient to be able to create pretty good results with only an audio recording, and these approaches will improve in further papers.

Lecture 21

2021-04-19T00:00:00+00:00

Paper Presentations IV

Thoughts

Frequency domain smoke guiding: while not especially physically motivated, this technique seemed quite useful for art directing smoke simulations - the ability to prototype in a low resolution (and hence real time) regime and have results actually translate to a higher resolution seems very useful. The differences in discretization are quite apparent when doing this without the guiding model, but with it the results are quite impressive, looking like simply more complex versions of the same phenomenon.
I thought the AnisoMPM paper was very interesting, although we didn’t dig too deeply into the math behind MPM. Their anisotropic damage models gave quite realistic results, however, as damage and deformations always followed lines of weaker material.

Lecture 20

2021-04-14T00:00:00+00:00

Deformables

Thoughts

Spent quite a bit of time talking about Monster Mash. The tool is quite fun and interactive, even for people with zero previous 3d modelling/animation experience. That much is impressive - I’ve done a bit of modelling/rigging/animation myself in the last year and I can say it’s very hard to get into, as I’m still quite bad at it. Their joint optimization for inflation works surprisingly well, with seemingly no issues on any inputs. I do wonder where it breaks down, and if it messes up with slightly disconnected/off drawings. Finally, it’s funny that no one had every just let the optimization run and render in real time in order to give the result a natural feeling of springiness and deformation. I think that actually made the user experience way better than it would have been if we had to wait for the optimization to converge every time.
I’m not too familiar with finite element techniques in general, but I imagine all the mechanical engineering topics like computing volumetric stress and strain are used in most fracture/physical deformation simulations. I wonder if elastic/plastic deformation are correctly modelled, and how we specify those for various (possibly anisotropic) materials. I have heard the Young’s modulus determines most of this, hmm…

Lecture 19

2021-04-12T00:00:00+00:00

Paper presentations III

Thoughts

The octree-based simulator was pretty impressive, although I’m a bit surprised nobody has made a (practical) implementation like this before. The discritization rules across octree resolutions seemed like pretty straightforward generalizations of the Eulerian uniform grid rules.
Ferrofluid simulation is super cool, and I would have thought you needed all the volumetric information in order to compute the magnetically coupled effects. The main insight (it seems) was that for fluids with zero viscosity, the surface velocity fully determines the interior - I guess this makes sense, since any changes can propagate instantly in order to make the interior a harmonic interpolation. The results were super impressive, and I could see this inspiring cool organic generation/art with their maze forming example.
I don’t think I understood all of the soap film dynamics paper, but it was cool that they compared results directly with real world phenomena and showed pretty compelling results. Were they rendering their soap films in a spectrally correct manner (leading to the color patterns), or was this more of a false-color curl visualization?
Similar thoughts for bubble rings & ink chandeliers paper - super compelling results when compared to the real thing (at least for bubble rings, ink chandeliers looked a bit worse). I wonder how they matched the initial conditions so closely. The math here was very complex, but the main idea of computing evolution in the primary configuration space of the filament was interesting.
Lastly, the snow paper. This one seemed pretty poplar, and gave very high performance results compared to other papers, but I didn’t think the results were actually that compelling - the snow looked too stiff and flat. I suppose that could be partly due to the opaque rendering, but the fracturing behavior also looked a bit more like soft styrofoam than snow. In any case, it is still pretty impressive to capture this behavior with relatively basic SPH rules, giving the high performance. I wonder if they can improve it to mach e.g. Disney’s snow from Frozen.

Lecture 18

2021-04-07T00:00:00+00:00

More fluids.

Thoughts

Most of the papers here were further optimizations of the basic Eulerian methods to better preserve energy and detail, as well as making fluid simulations faster.
I thought the PIC and FLIP methods were particularly interesting, as they combined the strengths of Eulerian and Lagrangian simulations by computing pressure and forces on a grid, but switching back to a particle based representation for advection. This allowed the non-linear advection term to correctly collect into new cells instead of some strange approximations seen last time.
Same thoughts about SPH as last time - it seems like the more flexible method overall, as it can be integrated with other types of simulation (rigid body, soft body, etc.), upgraded to do things like simulate snow and the material point method.