This paper contains many interesting ideas.  
The whole purpose is to have more realistic skinning   but make it suitable for interactive applications (i.e. fast enough, and with less memory usage).   
  
The paper starts by describing why linear blend skinning is not good enough (even though it is attractive because it is fast).  
  
The paper then describes the proposed scheme, which is based on the idea of having several input examples.  
There are some nice ideas which include  
1. adding more bones to the skeleton to overcome the shortcomings of linear blending. (any non linear system can be approximated by a stepwise linear system).  
2. adding more bones to simulate muscle bulges  

Finally, the paper gives the actual algorithm to compute the deformed skin given a new pose.  
The method starts by computing influence sets (which joints affect what), solving a bilinear problem  
to find weights and vertex positions (which is solved by successive approximation), and finally transforming normals to maintain correct lighting effects.  
  
The thing I like most about this paper is the idea of using more than one deformed mesh to do skinning.  
In my mind, I had the idea that we have to start from one deformed mesh and zero-pose, and we have to work our way  
starting from that. It is nice to see other approaches.