We use a Kinetics training/validation set with 239429/19715 videos. All the videos are extracted (utils_dir/ffmpeg_extract_frames.py) into frames and encoded with lmdb (utils_dir/make_lmdb_kin.py).
You may download UCF-101 dataset from UCF-101 and then extracted (utils_dir/ffmpeg_extract_frames.py) videos into frames
You may download HMDB-51 dataset from HMDB-51 and then extracted (utils_dir/ffmpeg_extract_frames.py) videos into frames