Kinetics-Data-Preprocessing

Kinetics-400 and Kinetics-600 are common video recognition datasets used by popular video understanding projects like SlowFast or PytorchVideo. However, their instruction of dataset preparation is too brief. Therefore, this project provides a more detailed instruction for Kinetics-400/-600 data preprocessing.

Download the raw videos

There are multiple ways to download the raw videos of Kinetics-400 and Kinetics-600. Here, I list two common choices that I found to be simple and fast:

1. Download the videos via the official scripts. However, I noticed that this option is very slow, so I personally recommend the next choice.

2. Download the compressed videos from the Common Visual Data Foundation Servers following the repository, which is much faster as they organized 650,000 independent video clips into several compressed files.

Resize the videos

The common data preprocessing of Kinetics requires all videos to be resized to the short edge size of 256. Therefore, I use the moviepy package to do so. The package can be easily installed by the following command:

pip install moviepy

Then, you can use the resize_video.py to resize all the videos within the given folder by following command:

python resize_video.py --size 256 --path YOUR_VIDEO_CONTAINER

IMPORTANT! Note that the resize_video.py will replace the original mp4 files. If you want to keep the original files, please make copys before resizing.

Prepare the csv annotation files

Following SlowFast, we also need to prepare the csv annotation files for training, validation, and testing set as train.csv, val.csv, test.csv. The format of the csv file is:

path_to_video_1 label_1
path_to_video_2 label_2
path_to_video_3 label_3
...
path_to_video_N label_N

The original annotations can be found at the kinetics website, or you can directly use download links of kinetics-400 annotations and kinetics-600 annotations. The official annotations support two different types of files: csv and json. However, both of them don't meet the above format. Therefore, I also provide a python code to transfer json files to the corresponding csv files with correct format. It takes two inputs: the container path of all videos, the path of official json annotation files. The output annotations will be named as 'output_XXX.csv' and located at the same folder. The label-to-id mapping dictionary will be saved as 'label2id.json'. The following command is my example.

python kinetics_annotation.py --train_path /home/kaihua/datasets/kinetics-train/ \
    --test_path /home/kaihua/datasets/kinetics-test/ \
    --val_path /home/kaihua/datasets/kinetics-val/ \
    --anno_path /home/kaihua/datasets/kinetics400-anno/