eduzhai > Applied Sciences > Engineering >

Dance Revolution Long-Term Dance Generation with Music via Curriculum Learning

  • Save

... pages left unread,continue reading

Document pages: 14 pages

Abstract: Dancing to music is one of human s innate abilities since ancient times. Inmachine learning research, however, synthesizing dance movements from music isa challenging problem. Recently, researchers synthesize human motion sequencesthrough autoregressive models like recurrent neural network (RNN). Such anapproach often generates short sequences due to an accumulation of predictionerrors that are fed back into the neural network. This problem becomes evenmore severe in the long motion sequence generation. Besides, the consistencybetween dance and music in terms of style, rhythm and beat is yet to be takeninto account during modeling. In this paper, we formalize the music-conditioneddance generation as a sequence-to-sequence learning problem and devise a novelseq2seq architecture to efficiently process long sequences of music featuresand capture the fine-grained correspondence between music and dance.Furthermore, we propose a novel curriculum learning strategy to alleviate erroraccumulation of autoregressive models in long motion sequence generation, whichgently changes the training process from a fully guided teacher-forcing schemeusing the previous ground-truth movements, towards a less guided autoregressivescheme mostly using the generated movements instead. Extensive experiments showthat our approach significantly outperforms the existing state-of-the-arts onautomatic metrics and human evaluation. We also make a demo video todemonstrate the superior performance of our proposed approach atthis https URL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×