WRITTEN COMPONENT ENTRY Style Translation [intial notes].
- style translation is the process of transforming an input motion into a new style while preserving its original content.
- Our solution learns to translate by analysing differences between performances of the same content in input and output styles.
- Style is a vital component of character animation. In the context of human speech, the delivery of a phrase greatly affects its meaning.
- Basic actions such as locomotion, the difference between a graceful strut and a defeated limp has a large impact on the tone of the final animation.
- Applications of human animation often require large data sets that contain many possible combinations of actions and styles.
- A database of normal locomotion, could be translated into crouching and limping styles while retaining subtle content variations such as turns and pauses. Our system can be used to extend the capabilities of techniques that rearrange mocap clips to generate novel content.
- Two motions in different styles typically contain very different purposes.
- Our model doesn’t account for kinematic interaction with the environment, the raw output may contain small visual artifacts, such as feet sliding on the ground.
- Such models (as in the paper’s) often require explicit frame correspondences to be solved during the translation process.
- Our solution draws its main inspiration from retargeting methods, which preserve content, but transform motion to adjust for skeleton size, kinematic constraints, or physics.
- Style is often difficult to encode in the procedural forms required by these methods.
- Statistical analysis of input-output relationships is a fundamental tool in nearly all areas of science and engineering. In computer graphics, such relationships have been used to translate styles of line drawings, complete image analogies and so on. Out method is the result of the comprehensive application of these principles to style translation.
- Blending and morphing approaches extrapolate from motion sequences with linear combinations.
- Such techniques allow for flexible synthesis of stylised motion, often with intuitive and tuneable parameters.
- Parametric approaches postulate a generative model of motion.
- Such general models of content would be needed to enhance a novice dance performance by removing mistakes.
- In all our evaluations, we acquired data in a mocap lab and processed it with standard commercial tools.
- Our data set contained various stylised locomotion’s: normal, limp, crouch, sneak, flap, proud, march, catwalk and inverse. Each style was performed at three speeds; slow, medium and fast.
- We found it took few (less than ten) to approach a near-optimal solution and a few more for full convergence.
- Style translation can operate on complex inputs; however, the translation itself must be relatively simple.
- Our system is intended for applications in which motions are specified by performance.
- It was not our objective to supplant existing motion synthesis techniques, but rather to present concepts that could be integrated into past and future work.
E. Hsu, K. Pulli, J. Popovic. “Style Translation for Human Motion”. Massachesetts Institute of Technology, Nokia Research Center.
No comments:
Post a Comment