Automated Extraction and Parameterizaton of Motions in Large Data Sets
------------------------
 
Another paper about motion blending! This paper takes the approach  
of the registration curves paper a little bit further.  
Now, we are not only interested in blending motions. We are also  
interested in "FINDING" the most suitable motions to blend.  

Once we find families of motions that are blendable. We can  
use them to create families of motions. We can blend all the motions  
in the family to create motions for movements in between the given motions.  

In the heart of the searching algorithm is the same Dynamic Programming (DP)  
 algorithm used to find the motion warping curve. However, we use  
it in a different way this time.  
The difference here is that we are now "comparing" 2 motions rather than just blending.  
Here, we are checking if it is ok to blend those 2 motions. The aturhos note  
that in order for 2 motions to be blendable, the best motion warping  
path must lie in the vicinity of local minima points in each direction.  

We can also use this to find only a range of matching frames between 2 motions (it  
doesn't have to be the entire motion). To do this, we apply the same  
DP algorithm to find the longest sequences that satisfies our  
conditions between the 2 motions.  If 2 of those sequences are really close,  
we can join them by a bridge.  Most probably, the user will not notice  
any difference, because the bridge is really short, and this way  
we can get longer matching motions.  

  
Now, we are ready to proceed. We can compare any 2 motions and find  
the matching parts in those motions. We can use that to build a graph  
that basically summarises all the possible matchings between  
motions.  
We precompute this graph, and we note that some of the motions that  
were un-blendable may now be blendable (transitive property).  

Once we have the graph of matching motions, we can now blend motions  
to create the families of matching motions. The paper gives some notes on  
how to select the best motions to start with to get good results, and  
avoid needless time overhead.