Date of Award
Honors Thesis (Colby Access Only)
Colby College. Computer Science Dept.
Dance improvisation, i.e. spontaneously generating and performing dance movements as the music plays, without prior choreography, is a challenging task due to its complexity and ambiguity. Previous works have shown inspiring results in tackling this problem with the combination of convolutional auto-encoders and generative autoregression. However, these works can only take on one dance genre at a time, which requires their training dataset to be hand-labeled. No work so far has exhibited meaningful results in synthesizing multiple dance genres with one single network system. In this paper, we propose an enhanced neural network architecture that features probabilistic motion generation. We train our network on our own dataset which contains dance clips of various genres, ranging from American street dance to Japanese Nico-Nico. Experimental results show that with the probabilistic setup, our system is able to reconcile the ambiguities present in the multi-genre dataset.
Machine Learning, Robotics, Audio Processing
Recommended CitationDeng, Yitong, "Data-Driven Automatic Dance Improvisation in 2D" (2020). Honors Theses. Paper 975.