Synthetic Pose Dataset

This thesis explores approaches to synthetic human pose generation, addressing the critical need for diverse and domain-specific datasets to train and enhance models for pose teaching software. With the ever-increasing reliance on machine learning methods that demand extensive training data, the development of synthetic datasets opens up new avenues for tasks such as personalised feedback or pose classification. I investigate two distinct methods for synthetic data creation within the yoga domain. The first approach, data augmentation, employs predefined rules applied to existing motion capture data. This method enables the precise definition of dataset variations according to specific task requirements. The rules applied during generation not only yield plausible poses but also provide valuable labels for training models. The second approach involves training a variational autoencoder, referred to as VAEGen, which generates new synthetic poses by sampling from a latent space and decoding them into pose representations. This method results in more diverse synthetic poses. Furthermore, the latent space representation offers numerous possibilities, including smooth pose transitions and the potential for synthetic movement generation. The application of synthetic data to train models for providing user feedback reveals promising results, with the system accurately predicting the rules governing the difference between two poses, achieving an accuracy rate of 72%. This capability enables valuable feedback to users, aiding them in aligning their poses more consistently with instructorguided positions.