-
Download raw videos from official website. The data structure should be like below:
kinetics/ ├── raw_video/ │ ├── abseiling/ │ │ ├──0347ZoDXyP0_000095_000105.mp4 │ │ ├──035LtPeUFTE_000085_000095.mp4 │ │ ├──...... │ ├── air_druming/ │ │ ├──03V2idM7_KY_000003_000013.mp4 │ │ ├──0aidJ-1R7Ds_000003_000013.mp4 │ │ ├──...... │ ├──......
-
Extract raw frames:
python3 dataset/process_data/create_lmdb.py --root-path data/kinetics/raw_video --dst-path data/kinetics/rgb_lmdb --num-workers -1 --data-type rgb --dataset kinetics
-
Extract optical flows:
python3 dataset/process_data/create_lmdb.py --root-path data/kinetics/rgb_lmdb --dst-path data/kinetics/flow_lmdb --num-workers -1 --data-type flow --dataset kinetics
-
Extract motion magnitude for each frame:
python3 dataset/process_data/create_lmdb.py --root-path data/kinetics/flow_lmdb --dst-path data/kinetics/flow_mag_lmdb --num-workers -1 --data-type mag --dataset kinetics
-
Extract motion magnitude for each clip
python3 dataset/process_data/create_lmdb.py --root-path data/kinetics/flow_lmdb --dst-path data/kinetics/ --num-workers -1 --data-type clip-mag --dataset kinetics
-
The final data structure should like this:
kinetics/ ├── raw_video/ ├── ├── ...... ├── rgb_lmdb/ ├── ├── ...... ├── flow_lmdb/ ├── ├── ...... ├── flow_mag_lmdb/ ├── ├── ...... ├── video_clip_mag_kinetics.pickle
-
Download raw videos from the official website. The data structure should be like below:
ucf101/ ├── raw_video/ │ ├── ApplyEyeMakeup/ │ │ ├──v_ApplyEyeMakeup_g01_c01.avi │ │ ├──v_ApplyEyeMakeup_g01_c02.avi │ │ ├──...... │ ├── ApplyLipstick/ │ │ ├──v_ApplyLipstick_g01_c01.avi │ │ ├──v_ApplyLipstick_g01_c02.avi │ │ ├──...... │ ├──......
-
Then you can prepare data for UCF101 with same procedure above and change the
--dataset
toucf101
. The final data structure of UCF101 should like this:ucf101/ ├── raw_video/ ├── ├── ...... ├── rgb_lmdb/ ├── ├── ...... ├── flow_lmdb/ ├── ├── ...... ├── flow_mag_lmdb/ ├── ├── ...... ├── video_clip_mag_ucf101.pickle