Producing "Open-Style" Choreography for K-Pop Music with Deep Learning

The goal of this thesis to generate natural and beat-matching choreography from music using deep learning. We compare different pose estimators to create a dataset of human figures for dance generation. Our deep learning framework comprises of a music encoder to create music features which is fed to a pose generator to create dance pose sequences and a music feature generator which reconstructed music features from output poses to improve music feature encoding. Our results showed a pose estimator with a GRU music encoder, generated more natural dance movements which match K-Pop music compared to previous work.