There are various editing or synthesis schemes to modify captured data for different purposes. Gleicher [6] proposed a method called “retargeting” to map a motion from the original skeleton to another one but keep significant constraints, as shown in Figure 4. For a walking motion, the feet must step right “on” the ground but not penetrate it or float up, and each footstep also needs to locate on the same spot.
(a) (b) (c)
Figure 4: Motion retargeting. (a) Original motion. (b) Retargeting to a smaller-scale skeleton.
(c) Retargeting to a longer body with shorter limbs.
Another mechanism is called “transition”. This technique smoothly connects the end of one motion clip to the start of another in time domain to generate a longer continuous motion.
In practice, transition does not always happen on either side of a motion but may take place on an arbitrary frame in a motion clip. The critical issue is to decide the best transition point which results in a smooth transition. It’s reasonable to assume that the more similar the two frames are, the more seamless transition becomes. Kovar et al. [13] proposed a distance metric to quantify the “similarity” of two poses and using a rigid transformation function to adjust one pose, so it will become the closest match for the target one. Then a graph structure was designed to record the transition nodes and motions that could be smoothly transited.
Further application falls on applying transition to a desired purpose like path fitting,
locomoting on a designed path, as shown in Figure 5. Gleicher et al. [7] rearranged the graph structure by collecting similar nodes into a hub node, and varied the possibility of transition.
The additional process to accomplish transition could be easily performed by applying transformations among the entrance, exit, and the “common pose”, which is the representative of the hub node.
Figure 5: Motion transitions apply to path fitting. The motions are generated to fit a word
“motion”.
To attain a smooth transition, an intuitive interpolation scheme is using a weight set of (0.5, 0.5) on both frames of the two motions at the transition point. Moreover, if we expand this interpolation to a few more frames, and apply a weight function from (1, 0) to (0, 1) gradually, we will get a smoother result. While extending this idea to the whole motion clip, the procedure becomes the foundation of “blending”. The main issue now turns to not only find the most similar poses of two motions but to figure out the total alignment of multiple motion clips.
Kovar and Gleicher [11] employed the distance cost metric proposed earlier on each pair of motions to form a 2D grid graph. This graph shows the similarity relations between all pair of frames of the two motions. Then, they find a minimal-cost connecting path that best aligns these two motions in time domain, as shown in Figure 6. After all paths were
combined together, blending can be applied to make new motions.
Figure 6: Time alignment path. Play motions according to the indexes of the alignment path makes these two motions do actions simultaneously.
They [12] use this idea to classify motions into different categories. Starting with a random sample motion, and apply the method to search other similar motions in the database as the outcome of classification. They also proposed an approach to directly control the blended motions. To achieve this goal, they defined and extracted parameters from a set of motion that best represent the type of those motions. For a punching motion set, the wrist joint might be the defined parameter, and the punching positions were extracted. For a walking cycle motion set, the root joint might be the defined parameter, and the vertical projections of final root positions were extracted. After denser sampling, more parameters will be produced, as shown in Figure 7. Collecting these parameters together forms a discrete parameterized motion space. Users can simply assign a location in the space, and the system will respond a set of blending weights that blend the sample motions most closely to the assigned location.
(a) (b)
Figure 7: (a) Before sampling. There is a significant error between the desired location and the result motion. (b) After denser sampling. More parameters are produced, and more accuracy is provided.
Other research uses various approaches to achieve motion interpolation. Grochow et al.
[8] presented an Inverse Kinematics (IK) system and a learning mechanism for motion data training. New motions can be edited by controlling a few constraint parameters. Mukai and Kuriyama [14] proposed a statistical approach to motion prediction and a modified Radial Basis Function (RBF)-like distribution function to construct a smooth parameter surface. And provided a way to estimate the reliability of the predicted motion, as shown in Figure 8.
Glardon et al. [5] translated different speeds and types of locomotion data to a new space and applied hierarchical Principal Component Analysis (PCA) for data dimension reduction and style classification. In the hierarchical PC structure, they discovered the relationship between the principal component and its corresponding speed value. Therefore, interpolation as well as little extrapolation of the speed value can be performed on this peculiar space to get new PC coefficient, so as to get the edited motion.
Figure 8: Reliability map. The blue areas represent high reliability, while the red areas represent low reliability.
The concept of motion editing by segmenting human body into parts and combining different parts from different motions was used by Heck et al. [9]. They treat human motion as a combination of upper-body action and lower-body locomotion. So the body is divided into two parts. They time-align the two target motions, and compute a rotation transformation for the splice point to keep natural balance. For a walking motion, when left arm swings forward, right leg should also be in front of the main body. In Figure 9, when a person carries a heavy box using both hands, the upper-body should leans back to balance the weight of the box.
Figure 9: When carrying a heavy box, the green motion leans back to balance the weight of the box.