This paper presents a method for searching a large database of motions given a reference motion. In addition, they present a method for constructing a continuous, parameterized space of motions by blending those search results that are nearby in parameter space.

Their search uses three key ideas: 1) by first searching for numerically similar motions and then using those as a basis for a broader search, they can find motions that are logically, but maybe not numerically similar. 2) Analyzing time correspondences is a good, simple way to determine numerical similarity. 3) Achieving interactive search rates can be made possible by first precomputing a match web, which they describe as an efficiently searchable representation of all possibly similar motion segments.

Their match webs are constructed for each pair of motion segments in the database by first constructing chains of minimum distance -- in a manner similar to their last paper, but extended to find *all* chains, not just the best. They then "bridge" the end of some chains to end up with a network of motion paths, representing all potentially similar pairs of motion segments.