Motion style retargeting

Michel Abdul Massih Said, Purdue University


Animation of a three dimensional character is the process of posing its figure at consecutive time steps, displayed at a specific speed to create the illusion of motion. Modifying an animation can be challenging once it has been created, and the process generally requires a long time to achieve. However, editing existing animations is often needed, consequently, there exists a demand for tools that assist animators to achieve such tasks with greater ease. This work focuses on a particular type of motion editing operation, which aims to extract the stylistic features from a source character and transfer them to a morphologically different target character. The method proposed in this research applies motion features extracted from motion capture data to user-specified groups of body parts on a target non-humanoid character. These body parts can range from a single skeleton joint to the whole character's body. The features applied to the non-humanoid character body parts are first extracted by a user-guided process from a motion captured humanoid character. This process begins through the identification of stylistic features on a source character by the user, who matches groups of body parts on the source and the non-humanoid target character. Motion features are then extracted from the source character, and finally transferred to the target character. This process is repeated for each of the body parts desired. A fast and intuitive retiming mechanism is also presented that allows the user to change the timing of the resulting motion as a post-process step. Once the motion style is extracted from the source character, the features can be saved and reused on any other character in the future. The main contribution of this work is the introduction of style transfer between characters with different morphologies by means of composing the style on the target character through the individual motion of groups of body parts. A number of examples of animations created using the proposed framework are presented, as well as results of a survey applied to a random sample of participants to categorize the style of the motions, rate their naturalness, and how well the animations express the motion styles.




Benes, Purdue University.

Subject Area

Computer science

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server