online action detection (OAD) is challenging since 1) robust yet
computationally expensive features cannot be straightforwardly used due to the
real-time processing requirements and 2) the localization and classification of
actions have to be performed even before they are fully observ