k400.md

简体中文 | English

Kinetics-400 Data Preparation

This document mainly introduces the preparation process of Kinetics-400(K400) dataset. It mainly includes three parts: Data Download, Extract the frames of the video file, Generate the path list for the frames and video files.

---

1. Data Download

For more information about the Kinetics, you can refer to its official website Kinetics; For the dataset download, you can refer to the ActivityNet ，which provides a download script. Furthermore,considering the difficulty in downloading K400 dataset, we upload the video files to Baidu cloud disk in the form of zip packages, and users can download it by themselves according to their needs. The total dataset needs about 135G storage space. Link ^{coming soon}

---

2. Extract the frames of the video file

To speed up the network training, we first extract frames from video files. Compared to the way of training through video files, training by frames could accelerate the training process significantly.

You can use the following command to extract the frames of video files.

python extract_rawframes.py ./videos/ ./rawframes/ --level 2 --ext mp4

After the video frames are extracted, they will be stored in the path of ./rawframes, which can be specified by the user, with a size of about 2T.

---

3. Generate the path list for the frames and video files

The path List file of video and frames of dataset will be provided to users directly through Baidu cloud disk，Link ^{coming soon}