Generating 3D Dances From Music Using Deep Neural Networks

This thesis focuses on the generation of original and unique 3D dances given a music
using deep neural networks. A state of the art model (Dance Revolution) was adapted
to take as input 3D data. Then it was trained using the recently published AIST++
dataset. At the generation phase, the model is able to generate credible dances. This
was achieved by introducing a novel audio data augmentation technique that modifies the
harmonic content of a song without changing the rhythmic content. This method allowed
for an increase in the number of training epochs before the LSTM network converges to
a static pose. Additionally, a novel method to evaluate the coherence of the generated
dances with respect to the style of music is proposed. The comparison is based on key
dance moves that are identified using the matrix profile. Using this method to evaluate
the dances, it was found that the model generate coherent dances with respect to the
dominant styles of music in the dataset.