Okan Arikan
UC Berkeley
Thursday, May 5, 2005
1065 Kemper Hall
3 :10-4:00 p.m.
Refreshments/reception to follow in in 1131 Kemper
Computer graphics involves creating images or animations that are ultimately judged by human users. Therefore visual plausibility is often more important that the absolute physical or numerical accuracy. We can take advantage of this observation to design efficient algorithms for rendering and animation.
I will present our motion synthesis research for animating synthetic characters that are ubiquitous in computer games and TV/Film production. Creating a realistic character motion is challenging, because people are very sensitive to the details in motion (due to individual style as well as complex anatomical structure of the characters). Novel motions can be created by rearranging pieces of already known to be visually pleasing motions. This allows us to obtain high visual quality without necessarily the absolute physical validity. The synthesis problem can be formulated as a combinatorial search, which tries to find an optimal arrangement such that the synthesized motion looks visually pleasing and performs the motions that the user wants. I will present combinatorial search methods based around Monte Carlo Markov chains and randomized dynamic programming. These methods can synthesize visually pleasing character motions that can satisfy geometric constraints ("Go to a particular spot") or perform specific actions ("Walk and then run"). In this context, I will also introduce a tool based on support vector machines for recognizing actions from motions. \
The visible color of a surface depends on how much light (power) it receives and hence is a crucial computation in photorealistic rendering. An important observation is that people are more sensitive to frequency content than the absolute value of the received power. My talk will include our research on approximating this received power while maintaining frequency content in lighting. Using such perceptually motivated approximations, we can render realistic images of complex environments an order of magnitude faster than other methods, without creating visually objectionable artifacts.