Eva Schiffer

Interactive Control of Avatars Animated With Human Motion Data
by Jehee Lee et al. SIGGRAPH 2002 

Summary:

This paper tackles the problem of controlling avatars in 3D environments in real time using two level representation of the motion generated semi-automatically from motion capture data. Special focus is given to the slightly different kinds of motion they attempted to create with their data sets and the types of user interfaces that they implemented.

Problem:

The authors' essential goal was to create a data structure from motion capture data that would allow for flexible and fast creation of motion data in order to move an avatar according to the input of a user in real time in a complex 3D environment. The authors' example implementations also touched on the sub-problems of generating good motion while path finding, navigating uneven terrain, navigating complex obstacles, and performing complex object interaction. 

Method:

The main method used in this paper was to process the motion capture data before hand to generate the two level data representation made up of a lower layer of information on the specific motions recorded and a higher layer of information about the clusters of similar motion and poses that were present in the data. At run time, the full space in the lower layer or the clustered representation could be searched to find good "next steps" for the animation. In the lower layer, good next steps were determined by the Markov model information that was previously collected and stored in the motion graph. In the higher layer, the clustering process embedded similar information about the probabilities of transitions between the clusters.

Key Ideas:

Motion capture data can be processed into a useful graph form which can be used to power avatar motion in real time in a 3D environment.
Different assumptions about which coordinate system to use when defining the distance of the root position and orientation can provide more power for different sorts of applications.
Clustering can provide a smaller and more abstract view of the motion graph which is easier to search at runtime, but may not be appropriate for all representations of the root position and orientation.

Contributions:

This paper provided a powerful way to generate motion graphs for producing avatar motion in real time from semi-arbitrary user input. It also provided very interesting applications of this technique to several sub-problems as well as interesting notes about the suitability of several user interfaces. 

Questions:

I would be interested to see how well this method actually handled constraints and fine motions, especially when using clustering. I am also curious to know if their system really scales up as well as they claim. It seemed to me that without the clusters things would devolve quickly, especially if their data contained many very similar motions.