Research Abstracts - 2006

Human Motion Comprehension

Eugene Hsu & Jovan Popovic

The high level goal of this graduate research is to give computers the ability to understand human motion. The simulated characters that we often see in feature films are the product of countless hours of work by skilled animators. However, the computers that generate this fantastic imagery have no notion of what the characters are actually doing unless the concepts are explicitly encoded through human intervention. This fundamental lack of comprehension makes it difficult to interpret and generate novel motions.

In prior work, we have enhanced the motion comprehension and synthesis abilities of computers through various decompositions of human motion. Our work on example-based motion control modeled implicit relationships between motion signals through semantically relevant decompositions [1]. For instance, in partner dance, a leader directs the motion of a follower through a series of beat-synchronized motion primitives. By exploiting the structure of dance, the work was able to synthesize an artificial partner for a motion-captured leader.

Our later work on style translation presented a motion filtering technique that changed the style of a performance [2]. It performed this task by using carefully chosen representations and statistical models that focused on encoding stylistic variations. The technique successfully changed the style of motions while preserving the intentions and nuances of the original content.

Our current research focuses on semantic representations of human motions. Its goal is to construct mappings between intuitive descriptions of actions and low-level motion signals. Doing so enables a number of interesting applications. One could map signals to action descriptions to construct annotations, which could then be used to drive user interfaces, compress motion signals, or perform textual summarization. The ability to synthesize motions from descriptions enables rapid scripting of animations for both offline and interactive applications.

At the highest level, this work aims to empower content authors by allowing them to manipulate visual data using intuitive controls. Although this project focuses on applications in human motion synthesis, the techniques that will be developed may also be applied towards other areas within computer graphics.

References

[1] Eugene Hsu, Sommer Gentry, and Jovan Popovic. Example-based control of human motion. Proceedings of the 2004 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, pp. 69-77. 2004.

[2] Eugene Hsu, Kari Pulli, and Jovan Popovic. Style translation for human motion. ACM Transactions on Graphics 24(3), pp. 1082-1089. 2005.