Existing methods for parameterization of motions are tailored more for small data sets with a restrictive range of variation.  This paper provides a method for identifying logically similar motions from a large data set and building a continuous and intuitively parameterized space of motions from those similar motions.  Topics discussed in the paper include how to search motion data sets for similar motions and a method for creating new parameterized motions from example motions.

Key ideas:

1. In order to find logically similar motions, the search query first finds numerically similar motions and then broadens the search by using these closer motions as part of a new search query to help find more distant motions.  Numerical closeness is determined by identifying corresponding frames and comparing their average distance against some threshold value.

2. Before starting the search, a match web is computed, which is a compact and efficiently searchable representation of all possibly similar motion segments.

3. Parameterized motions are generated by blending motions and picking out relevant features from these motions by using a user-specified parameterization function.  From this function, one can calculate an inverse that will dispense a set of blend weights that produce the appropriate motion.


Contributions: For searching motion data sets, the paper contributes the ideas of doing a multi-step search, using time correspondences to determine similarity, and interactivity through precomputation of a match web.