Stylistic Generation of Conducting Motions from Examples

We come into contact with animated or virtual humans more often than we think: in movies or computer games for example. The motions which these characters perform appear realistic to us and are often the result of a combination of motions that have been recorded from “real” humans (Motion Capture) and animations that have been carefully designed by experienced professionals. This process is both expensive and time consuming. Once the animations have been created they are fixed, making it difficult for a digital character to adapt its movements to changes in the surroundings. As an alternative, a computer can be employed to generate movements based on a specification of abstract goals. Generated movements can be very efficient, however they often lack realism and have a “robotic” feel to them. This research focuses on the combination of the realism gained from recorded movements combined with the speed and flexibility of computer generated motion sequences. The possibility to capture a person’s characteristic “style” of movement from a limited set of recorded example motion sequences is explored as well the possibility of using this information to generate unique movements that reflect a person’s style. A number of “style parameters” are defined that capture elements of a person’s movement style, based on a review of available literature on the subject. A newly developed software framework is employed to analyze pre-recorded example motions and calculate values for these parameters, which at a later stage are used in an attempt to generate original motion sequences in that motion style. The differences between the values that are found for the motion style parameters are tested for significance, in order to determine which parameters most successfully distinguish between different test subjects. In order to evaluate the quality of generated motion sequences, a short study has been performed in which respondents are asked to identify people based on example generated motion sequences. The results of this study indicates that the success rate of identifying a person is close to random, suggesting that the current implementation does not yield motion sequences that realistically mimic the style of a human. For the evaluation of the methods presented here, conducting gestures have been used as a sample motion type. Conducting motions must adhere to certain rules, making it possible to generate them using software, however there is a great amount of freedom for the conductor to introduce nuances in the actual movements that are performed.