Laplacian coordinates-based motion transition for data-driven motion synthesis

As a fundamental technique in data-driven (or example-based) methods, motion blending has been employed to produce new motion clip from two or more clips or introduced as a medium to improve the final result of other method, for example, motion graphs. As methods of motion blending always require the system or users to classify the motion samples, it is difficult to accomplish motion blending automatically and to preserve local details, such as angular accelerations of joints of original motions. In this study, the authors developed a framework based on Laplacian coordinates to produce transitions between two motion clips, emphasising on preserving local details, without any need of motion classification. At first, multidimensional Laplacian coordinates are introduced to present local details of joints´ angle. Then an error function is deduced to measure how much result motions differ from original ones in local details. By minimising such error function, new motions can then finally generated, which can preserve local details of input motion clips as much as possible. The following experiments show the effectiveness of the authors´ framework.