简体中文 | English
Kinetics-400 is a commonly used benchmark dataset in the video field. Please refer to its official website Kinetics for details. You can refer to the official address ActivityNet, and use the download script provided to download the dataset.
Considering the difficulty of downloading the K400 data set, we provide two download methods: (1) Baidu network disk download (2) Script download
Netdisk link: https://pan.baidu.com/s/1S_CGBjWOUAuxL_cCX5kMPg
Extraction code: ppvi
- Download the training set link list file train_link.list and the validation set link list file val_link.list.
Write the download script download.sh as follows:
file=$1
while read line
do
wget "$line"
done <$fileDownload training set command:
bash download.sh train_link.listDownload verification set command:
bash download.sh val_link.list| category | Number of data | list file |
|---|---|---|
| Training set | 234619 | train.list |
| Validation set | 19761 | val.list |
-
After downloading, unzip and add the data path to list file.
-
Due to the failure of some video link, part of original data is missing. This copies need about 135G of storage space.
This copies is only used for academic research. If it is helpful to you, welcome to star our project
In order to speed up the training process of the network, we first extract frames from the video file (K400 video file is in mp4 format). Compared with the method of network training directly through video files, the method of frames can greatly accelerate the speed of network training。
Enter the following command to extract the frames of the K400 video file
python extract_rawframes.py ./videos/ ./rawframes/ --level 2 --ext mp4After the video file frames are extracted, they will be stored in the specified ./rawframes path, and the size is about 2T.
| category | Number of data | list file |
|---|---|---|
| Training set | 234619 | train_frames.list |
| Validation set | 19761 | val_frames.list |