Evaluating Data Driven Character Animation

Humanlike characters are an important area of computer animation. These characters are vital components of many applications, such as computer games, cinematic special effects, virtual reality training, and artistic expression. There are two main challenges to animating humanlike characters. First, our daily familiarity with human motion makes creating realistic humanlike motion particularly difficult, since viewers have an extremely high sensitivity to artifacts or errors in human motion; this is the problem of producing animation of sufficient quality. Second, animators require a motion generation system that can produce all motions required to animate the character through the full range of tasks in the target application; this is the problem of ensuring the motion generation system has sufficient capability. While several technologies have been developed to address these problems, our understanding of such technologies is still developing. Accordingly, techniques for producing high-quality animations—such as motion capture— and for producing high-capability animation systems—such as motion graphs— typically rely heavily on the intuition and abilities of a skilled animator to achieve acceptable results, and hence are difficult to use or automate. The goal of this work was to augment the animator’s intuition by providing ways to more objectively quantify and measure a system’s capability to produce the required motion, and the quality to viewers of the motion so produced. We present a technique for evaluating motion quality by measuring user sensitivity to editing-based artifacts in captured motion, derive practical animation guidelines from the results, and discuss several interesting systematic trends we have uncovered in the experimental data. We also present an efficient technique for evaluating the capability of a motion graph to fulfill the requirements of a given scenario, along with an examination of capability deficiencies it uncovers and the effect on those deficiencies of some methods commonly used to improve motion graphs. Finally, we propose a wide range of extensions and applications enabled by these new evaluation tools, such as automatic optimization and vetting of motion graphs.