For Kinetics, we decode on the fly, each row in the txt file include:
video_path class
And we load the videos directly, please place the training set in SSD for fast IO.
Prepare dataset (UCF101/diving48/sth/hmdb51/actor-hmdb51), and each row of txt is as below:
video_path class frames_num
These datasets saved in frames. We offer list for all these datasets.
Download split datasets lists from here Google Driver
mkdir datasets
cp -r [path_to_lists] datasets/
mkdir experiments && cd experiments
mkdir logs
As some Youtube Link is lost, we use a copy of kinetics-400 from Non-local, the training set is 234643 videos now and the val set is 19761 now. All the videos in mpeg/avi format.
These two video datasets contain three train/test splits. Down UCF101 from UCF101 and HMDB51 from HMDB51.
Down Sth-v1 from Sth-v1.
refer utils/data_process/gen_hmdb51_dir.py and utils/data_process/gen_hmdb51_frames.py for details.
We also provide manuscript to get the sub-set of kinetics-400. Please refer to Contrastive/utils/data_provess/semi_data_split.py for details.