1. Sparse Rig Parameter Optimization for Character Animation
J. Song, R. Riverba, K. Cho, M. You, J. P. Lewis, B. Choi, J, Noh
Johnno De Vos and Joris Brandt

2. Summary: What are we talking about?
A technology to help retarget motion from one rig to another with minimal loss and/or computational power
Example: record animation on non-final model, transfer it to new one later
Useful for retargeting motion capture data to a rigged model

3. What is Motion Retargeting?
It’s a process by which animation/movement is translated from one rig or mesh to another
Can be very challenging with very different target objects
Used a lot for motion capture data, used to rebind data from mocap skeleton to the final model
Photo credit: Edu.tr

4. What is Sparse Rig Parameter Optimization
Motion retargeting using an intermediate object
Animator has to do less manual editing
Transferring source motion to the intermediate mesh
Optimizing the intermediate object motion to reduce errors
Robust handling of various types of input

5. Input Takes source motion and rigged target character as input
Robust system that takes various input
Motion Capture
Skeletal motions
Mesh animation
Any combination

6. Skeletal Motion Retargeting
Intermediate skeleton is created based on source motion and target
If source and target have skeleton, intermediate skeleton is created
Source motion is retargeted to intermediate skeleton
Using Lie group representation saves costs
Lie group provides linearized manifold

7. Mesh Deformation Transfer
Take corresponding points on source and target respective surfaces
Create a mesh using Delaunay tetrahedralization of source points
Make intermediate mesh Uses mesh structure of previous created mesh Uses points of the target character
Transfer the source motion to intermediate object Transfer of deformation gradients
Delaunay tetrahedralization gebruikt cirkels om midden van driehoeken te bepalen en maakt daar een mesh van

8. Keyframe Extraction Source animations may contain too many frames
Multiscale motion saliency is measured for each frame Calculates most important frames
Extract keyframes that best describe the motion
Saves computation time and editing for animator
Motion saliency from Halit and Capin
Calculated by taking Gaussian weighted average of each point

9. Rig Space Optimization
Optimization is posed as a minimization problem
Trying to minimize the difference between intermediate and target character
Minimizing the energy difference in a frame
p(c) ≈ p(c(t))+Jit (c(t))W(c(t))(c-c(t))
P = pose C = space rig parameters T = frame Jit = Jacobian of target character rig at iteration step W = csontraint weight matrix (identity matrix if undefined)

10. Rig Space Optimization
Temporal Smoothness and lack of popping is ensured by limiting acceleration energy at a given time-step
For further optimization energy is one again limited. Motion errors for each keyframe are compared to a global motion error threshold. For each keyframe motion error count that’s too high a new keyframe is generated taking over the maximum pose error.
Prior Jacobian has necessary components updated at each iteration, if required.

11. Related Work: Optimizing motion in Rig Space
(Hahn et al. 2013)
Augmentation of existing Keyframe animation procedures: secondary motions generated through physics simulation of movement based on existing rig.
Allows keeping of source rig, which saves on work time.
Can be further edited by simply modifying existing parameters of the rig space

12. Related Work: Inverse Rig Mapping
(Holden, Saito, and Komura 2015)
Using non-linear regression on example animation provided. Learn the motions from the example data to estimate the motions on a different rig.
Similar goal: increase productivity by providing system which could remap animation out of the box.
Problem: data dependant, quality of results directly proportional to quality and consistency of input data. Does not generalize well to rigs with too many differences.

13. Results

14. Results
Two teams: one experienced, one slightly less experienced.Asked to make animation with one of three techniques: Keyframing, using commercial software, and using Sparse Rig Parameter Optimization
Animation
Length in Frames (30 FPS)
Working time
Sword Fight
1140
Keyframing
6.4 Hours
Mocap + Software
3 Hours
Mocap + SRPO
1.8 Hours
Characters Dancing
2700
6 Hours
2.1 Hours

15. Results Sample size very small: not necessarily representative
Generally less error-prone than other techniques
Sparse Rig Parameter Optimization reduced work time in both examples

16. Conclusion
Sparse Rig Parameter Optimization can be a valid technique for creating animations from motion capture data. With correct setup of rig parameters can cut down significantly on working time, and increase productivity.
More working examples required before determining if the technique is innately superior to other, commercially available systems and techniques

17. Thank you for listening! Any questions?

18. Discussion
Is the system’s generalisation actually good if it’s so dependent on inputs?
Is it useful for large animation teams if individual preferences or differences in style can generate errors?
Perhaps less so, if all time gained is spent making up for issues in production
Will this system improve or degrade productivity for smaller animation teams?
If it saves time for animators, probably improve