S-Space College of Engineering/Engineering Practice School (공과대학/대학원) Dept. of Computer Science and Engineering (컴퓨터공학부) Theses (Ph.D. / Sc.D._컴퓨터공학부)
Computational Approaches on Choreographing Multiple Actor Motion
컴퓨터를 활용한 여러 사람의 동작 연출
- 공과대학 전기·컴퓨터공학부
- Issue Date
- 서울대학교 대학원
- Graphics; Character Animation; Multiple Actor; Choreography; Author-ing; Physics-based Control; Deep Learning; Reinforcement Learning
- 학위논문 (박사)-- 서울대학교 대학원 공과대학 전기·컴퓨터공학부, 2017. 8. 이제희.
- Choreographing motion is the process of converting written stories or messages into the real movement of actors. In performances or movie, directors spend a consid-erable time and eﬀort because it is the primary factor that audiences concentrate. If multiple actors exist in the scene, choreography becomes more challenging. The fundamental diﬃculty is that the coordination between actors should precisely be ad-justed. Spatio-temporal coordination is the ﬁrst requirement that must be satisﬁed, and causality/mood are also another important coordinations. Directors use several assistant tools such as storyboards or roughly crafted 3D animations, which can visu-alize the ﬂow of movements, to organize ideas or to explain them to actors. However, it is diﬃcult to use the tools because artistry and considerable training eﬀort are required. It also doesn’t have ability to give any suggestions or feedbacks. Finally, the amount of manual labor increases exponentially as the number of actor increases.
In this thesis, we propose computational approaches on choreographing multiple actor motion. The ultimate goal is to enable novice users easily to generate motions of multiple actors without substantial eﬀort. We ﬁrst show an approach to generate motions for shadow theatre, where actors should carefully collaborate to achieve the same goal. The results are comparable to ones that are made by professional ac-tors. In the next, we present an interactive animation system for pre-visualization, where users exploits an intuitive graphical interface for scene description. Given a de-scription, the system can generate motions for the characters in the scene that match the description. Finally, we propose two controller designs (combining regression with trajectory optimization, evolutionary deep reinforcement learning) for physically sim-ulated actors, which guarantee physical validity of the resultant motions.