Style Graph for Human Motion

This paper presents a novel model called a style graph for generating style-controllable human motions.Styles are considered as the space-time transformations that project a single content into different behavioral subspaces.For an input human motion, style and content can be extracted via aproposed principal rotation analysis.Then a candidate style set is pruned according to the input style, and then a style graph can be automatically constructed from the remaining styles.Finally, users are allowed to control the style of the input motion with the style graph.We evaluate our model by controlling the styles of various human motions (such as dribble and airplane motions) with a single style graph that is constructed from a stylistic walking database.The synthesized motions own the characteristics of user-specific styles while preserving the contents of the input motions.Experiments demonstrate that style graph is effective for generating style-controllable human motions.