DANCING AI

Bachelor Thesis Project - Spring 2021, Tilburg University

Generating choreography is challenging as it requires knowledge in both dance techniques and musical elements. In this work, to generate a high-quality dance dataset, different pose estimation methods were explored, and a pose cleaning method was developed. For the actual dance generation task, an existing framework was modified to investigate different music encoding approaches and to improve the performance. Specifically, LSTM and GRU were tested as a music encoder, and a novel music feature generator was proposed to reconstruct music features from the generated dance. The results demonstrate that the proposed approach with GRU as a music encoder performs better than the existing model, creating natural dance movements matching with the music.