This paper discusses computer puppetry. The 2 major differences between just   
doing normal motion capture and computer puppetry are.
  

 1. Real time and online performance: The actors movements must   
be transferred to the virtual character in real time without delays.
  

 2. Retargetting: Virtual characters will, in general, not have   
the same proportions as the original actor.  
  
The paper addresses those 2 problems. One of the implications   
of the first problem is that we need faster algorithms.   
This is why the paper tries to find analytical solution to   
all problems because they are, in general, faster to compute.  
Another implication of the first problem is that we can't see   
the future. We have to do all the processing and output   
rendering immediately, so some of the algorithms that need to   
run offline will not be suitable here! This is why, for example,   
the paper uses Kalman filters for motion filtering. Kalman   
filtering runs online, as opposed to gaussian or median filters   
which require us to look into the future.  
  
The second problem is retargetting. Retargetting without losing   
anything is impossible. We can preserve joint orientations,   
or end positions, but not both. The paper describes an importance   
analysis method to find which way to go (clearly, doing anything   
in between, will result in a bad motion. You have to choose one   
way to go! You can only be inbetween if you are in a transition phase)  
  
Preserving joint positions opens another problem,   
which is choosing an IK solver to find the orientations for those   
positions. The paper describes an algorithm to do IK solving   
for the whole body in real time.