Dance Style Classification by LSTM RNN
Yanika Mifsud, Frankie Inguanez · 2021
This study falls under a big data problem: the aim to classify dance styles in an unbiased manner and generate accurate SEO tags automatically. Addressing a part of this content classification, led this research to classify 2 dance styles by solely pose estimation. OpenPose, a pose estimation generating library, was used to extract 18 key-body points from an online gathered dataset to create temporal motion sequences. Multi-class classification was implemented in this approach to recognise Ballet and Breakdance styled choreography. An LSTM recurrent neural network modelled the structure and numerous configurations were tested. The best model achieved an accuracy of 90% with a 38% loss. Insightful observations regarding the dataset’s collection criteria were noticed and recommended.