process_data

Process data

This folder has some tools to process UCF101, HMDB51 and Kinetics400 datasets.

1. Download

Download the videos from source: UCF101 source, HMDB51 source, Kinetics400 source.

Make sure datasets are stored as follows:

• UCF101

{your_path}/UCF101/videos/{action class}/{video name}.avi
{your_path}/UCF101/splits_classification/trainlist{01/02/03}.txt
{your_path}/UCF101/splits_classification/testlist{01/02/03}}.txt

• HMDB51

{your_path}/HMDB51/videos/{action class}/{video name}.avi
{your_path}/HMDB51/split/testTrainMulti_7030_splits/{action class}_test_split{1/2/3}.txt

• Kinetics400

{your_path}/Kinetics400/videos/train_split/{action class}/{video name}.mp4
{your_path}/Kinetics400/videos/val_split/{action class}/{video name}.mp4

Also keep the downloaded csv files, make sure you have:

{your_path}/Kinetics/kinetics_train/kinetics_train.csv
{your_path}/Kinetics/kinetics_val/kinetics_val.csv
{your_path}/Kinetics/kinetics_test/kinetics_test.csv

2. Extract frame and flow

[Optional for Kinetics400] Reduce the video dimension to short size = 256 pixel with ffmpeg, to save space and time. An example is provided in src/resize_video.py.

Edit path arguments in main_*() functions, and python extract_ff.py. Video frame and TV-L1 optical flow will be extracted.

3. Collect paths into csv

Edit path arguments in main_*() functions, and python write_csv.py. csv files will be stored in data/ directory.

then prepare data/ClassInd.txt that stores sorted action names for corresponding dataset, like

ApplyEyeMakeup
ApplyLipstick
Archery
...

is for UCF101.