This paper presents "automated methods for identifying logically similar motions that are numerically dissimilar" and automatic registration and application of blending techniques to "create" a continuous space of motions.


Key contributions

 * A method for searching motion data sets
 * Creation of new motions from the example motions using blending and parameterization of the space of relevant motions given a set of target parameters.
 
Motion search
  * "find 'close' matches that are numerically similar to the query and then use them as queries to find more distant matches"
  * Use of "match webs" - compute distance for each set of frames - find chains of 1D minima and connect nearby chains using "bridges"
  * Searching within match webs - iterative search algorithm until no new matches are found - build a match graph.

Parameterizing motion
 - blend example motions together to create new motions.
 - motions must be registered in time
 - sample subsets of blend weights "by finding sets of example motions that are nearby in parameter space"