The proposed framework provides a useful tool that assists artists to transfer rig parameters of an animation between different models. We used four artist-created models, Arhat, Nezha, FlourSack and CyclopsBoy, in our experiments. A two artist-created animations of Arhat were used as source animations. The numbers of controllers and parameters involved in our experiments are listed in Table 4.1.
There are often numerous controllers and parameters of a model. However, under the consideration of system efficiency, accuracy and the artist’s conventions, artist can choose the really relevant ones which may be used in this animation beforehand.
For example, as show in the Table 4.1. The number of parameters of Nezha is 603, but only 122 parameters should be involved in the running case.
As mentioned in Chapter 3, during the Motion Transfer step, our system firstly predict the desired target meshes according to the source motions. Next, the target rig parameters is recovered to fit the predicted mesh while preserving the undeforma-bility of model structure and the temporal smoothness of parameters in Pararmeter Recovery step. Finally, when artist refines the resultant animation, the previous steps are re-executed and these changes will be propagated throughout the whole animation. The results of each steps are shown in Figure 4.1
Figure 4.2 demonstrates the effects of using temporal smoothness in motion transfer. With the assistance of time coherence term, the target animation becomes smoother and closer to the source motions. In addition, the information of the
Table 4.1: The numbers of controllers and parameters used in experiments.
Type Model # of controllers # of parameters
Original Arhat 77 610
Nezha 85 603
Floursack 28 128
CyclopsBoy 33 253
Testing case Model # of used controllers # of used parameters
Running Floursack 4 35
CyclopsBoy 7 21
Nezha 21 122
Walking Nezha 21 73
user-specified first frame is clearly propagated to the following frames. Substantial improvements can be observed on the postures of the back and legs in the target motion.
Figure 4.3 demonstrates the influences of energy terms in rig parameter recovery.
Without the constraints of undeformability term Eu, the mapped mesh could be skewed badly as the skewed spine shown in the rightmost column of the second row.
Although the temporal coherence term Es slightly degrade the motion fidelity, it effectively maintains the stability of rig parameters which is important to animators for further editing. The forth row shows that the right shank controller (in a blue car shape) is better transferred with the use of Es. The temporal smoothness can also help maintain the model structure in some degree. For example, the spin is preserved by the temporal smoothness even though we do not use Eu in the forth row. Finally, the inclusion of all three energy terms achieves the best stability and performance as shown in the last row.
Figure 4.4 demonstrates the improvement of using motion refinement. After adding three keyframes, the back and leg postures can be clearly improved. Through the refinement, the over incline degree of model’s back is fixed and the bending posture of its left leg become more nature. In our experiments, users can obtain satisfied results by editing at most three to five keyframes with only a few operations.
It shows that the resultant rig parameters are editable and our proposed motion refinement process can effectively propagate user’s editing which reduce much user’s
effort.
Figure 4.5 shows another example of a walking animation. Since motions in this sequence are slower and more subtle, it is more difficult than the running animation.
For such challenging task, our method can still produce acceptable results.
To prove the ability of our framework to work smoothly with various models and different rigging systems, we use two extremely different artist-created models, Arhat and FlourSack, as our input. The artist-created running cycle animation of Arhat is used as source animation. As show in Figure 4.6. in spite of the large difference in rig structure and model topology, our system can still work well and produce a nice running cycle animation of FlourSack. Another case is shown in Figure 4.6. We test more case as show in Figure 4.7. In this example, we use Arhat and CyclopsBoy as our source and target models, and retarget the running animation from Arhat to CyclopsBoy.
Figure 4.1: Step results. Top row: source motions. Seccond row: results in motion transfer step. Third row: results in rig parameter recovery step. Fourth row: refined results in motion transfer step. Bottom row: refined results in rig parameter recovery step.
Figure 4.2: The comparison of adding temporal smoothness in motion transfer.
Top row: source motions. Middle row: original deformed results without temporal smoothness. Bottom row: the results with temporal smoothness.
Figure 4.3: The impacts of different energy terms in rig parameter recovery. In this example, we focus on the right shank controller (the one in a blue car shape) and the spine skeleton (the red line in the fourth column). Top row: source motions.
Second row: results using the energy function with only Em. Third row: results using the energy function with only Em and Eu. Forth row: results using the energy function with only Em and Es. Bottom row: results using the energy function with all three terms.
Figure 4.4: Motion refinement. Top row: source motions. Second row: results without refinement. Third row: refined results after editing three keyframes. The artificial animation created by artists is shown on the bottom row as a comparison reference. On the rightmost column, we show how the left leg is improved through user refinements.
Figure 4.5: Another motion retargeting example of the walking cycle animation.
Top row: source motions. Middle row: results without refinement. Bottom row:
refined results after editing three keyframes.
Figure 4.6: Motion retargeting example of vastly different models. In this case, we use two extremely different artist-created models, Arhat and FlourSack, as our input. Top row: source motions. Middle row: results.
Figure 4.7: Motion retargeting example of different rigging systems. Source model is Arhat and target model is CyclopsBoy. Top row: source motions. Middle row:
results without refinement. Bottom row: refined results after editing four keyframes.