The use of synthetic hand pose data in training sign language fingerspelling models

This thesis presents a deep learning model that was developed to recognise sign language fingerspelling sequences and that was trained entirely on computer generated training data. This important contribution to sign language recognition (SLR) confirms that the distributional shift between the synthetic and real world domains can be mitigated through the use of suitable data augmentation. Furthermore, the model architecture is significantly more efficient than those proposed in previous SLR works; it enhances its suitability as a real-time assistive application for a Deaf user.