SIGGRAPH 2003: Character Animation papers

The first paper session on Tuesday covered character animation of the human form. The papers in this section were quite good and may lead to some interesting technologies for folks such as Alias (formerly Alias|Wavefront) and others to consider in new products for topics such as synthetic dance and scripted motion using timeline painting.

Rhythmic Motion Synthesis through Sampling

This was a fascinating paper on using motion-captured data of rhythmic movements (dancing was the main example) in order to later synthesize similar movements using computer generated characters. The talk concentrated on two parts, the computer analysis of the MoCap data and the synthesis of the analyzed data into output.

Analysis was performed in order to determine which movements were unique and which ones were variations on existing movements. By separating these two types of data and creating transition graphs for each step, governed by the analyzed musical beat of the background, they were able to create random, credible animation of dancing individuals and pairs without significant human intervention.

The synthesis phase combined the stochastic generation with certain rules for proximity (to keep the dancers from running into each other) and direction.

I'm looking forward to reading the paper on this one.

Motion synthesis From Annotations

This group of researchers from Berkeley worked with a large (15,000 frame) data set provided by EA to create annotated groups of movements which could then be assembled in real time using a timeline painting technique.

Again, the analysis is done by computer (because 15,000 frames would take a long time to analyze), annotating motion data by using machine learning and human guidance. By offloading the work to the computer, they were able to classify such operations as walking, running, jogging, waving, etc., and use these annotations for choosing the next animation.

Unfortunately, the technique requires exhaustive MoCap data, because there is no interpolation between frames. However, if you have a large amount of data, it is an excellent real-time system.

Layerd Acting for Character Animation

These researchers presented a new mechanism for iteratively animating models in an intuitive fashion. By using video capture and (literally) tinker toys, they have created a system that allows you to "layer" new behaviors onto existing animations to create more complex animations. As such, beginners are able to create complex animations in a stepwise fashion by creating the general movement direction, then concentrating on the feet, then arms, then torso, etc.

Benefits for experienced animators are questionable, but the acting model presented here looks excellent for beginners and for other types of control (like camera controls).

Their lab environment had a large screen (it appeared to be 10'+ diagonal) displaying the current animation and mirroring the users actions along with optical sensors to track the tinker toys. They had clearly put a lot of thought into the mechanism, as there was support for varying users (your arms may be shorter than a colleague's, etc.) as well as defining arbitrary ground planes (just add a table of any height).