Addressing the data diversity gap with uniquely generated synthetic videos for real-world human action recognition