Eva Schiffer

Automated Extraction and Parameterization of Motions in Large Data Sets
by Lucas Kovar and Michael Gleicher. SIGGRAPH 2004.

Summary:

This paper presents a system for "identifying logically similar motions" and using these similarities to create match webs which describe the transitions between these motions in a more flexible continuous space of motions that can be expressed in a parameterized form unlike standard motion graphs.

Problem:

How does one define new motions based on existing motion capture data? How can one generalize motion capture data into a more controllable, parameterized form?

Method:

Given some measure of distance between two motions, a chart of difference between the example motions at different times is formed. In order to find the match web, first the valid regions in this graph are found. These valid regions are less restrictive than the corresponding paths in a registration curve's dynamic search. The valid regions correspond to times where motions are "similar enough" that we may be able to form useful blends. A threshold parameter can be applied to determine if significant valid regions even exist. Now using the distance grid in the numerically similar areas that we found with the valid regions, we can compute local minimal paths through time. Since these paths may not be sufficient to connect across the whole range of time/motions bridges can be added to improve connectivity. The set of all the minimal paths between all the numerically similar motions defines the match web for the motion space. Since all motions can be directly compared in the match web, it is simple to search for blends which will move between a particular starting point and ending point. 

Key Ideas:

Numerically similar motions from motion capture data can be used to define parameterized spaces of motions.
Because the match web is pre-computed, finding desired blends is quite efficient. 

Contributions:

The ability to find numerically similar motions.
The ability to pre-compute information about optimal blends between a large number of motions.
The ability to use motion capture data to create a parameterized representation of a continuous range of motions.

Questions:

I am no where near as clear as I would like on how one gets from the match web to the exact parameterization of a motion space.