Region-sequence based six-stream CNN features for general and fine-grained human action recognition in videos (2018)