Author(s): Jonah Hoff
Mentor(s): Prosenjit Chatterjee
Institution SUU
The recognition of ballet poses combined with feedback for alignment and adjustments can be crucial for dancers learning ballet. This project uses a neural network model that recognises and recunstruct’s ballet poses. The model is trained with data from the Let’s Dance(1) dataset, containing images paired with JSON annotations for key joint locations. We trained a model to recognise key joints and reconstruct them from user provided image inputs. The average error of the points and visual inspection(2) were used to verify the accuracy of the model’s ability to provide precise feedback. The model predicts key points with an average of a 55 pixel difference between the original points and the predicted points. This work demonstrates methods to apply automated feedback to the ballet learning process. Future works could expand the dataset and improve model accuracy when fed diverse poses.