Dataset Installation Instructions

November 11, 2024 · View on GitHub

We provide detailed instructions for download and setting up the Kinetics-400 and Something-Something V2 datasets.

Kinetics-400

To download the Kinetics-400 dataset, you can visit the following page: https://github.com/cvdfoundation/kinetics-dataset. The same process applies to Kinetics-600 and 700, but we only evaluated on and provide checkpoints for Kinetics-400.

We restate the instructions below:

  1. Clone the Kinetics-400 repository:

    git clone https://github.com/cvdfoundation/kinetics-dataset.git
cd kinetics-dataset

  2. Download the Kinetics-400 video tarballs:

    bash ./k400_downloader.sh

  3. Extract the Kinetics-400 video tarballs:

    bash ./k400_extractor.sh

  4. Download the Kinetics-400 annotations from the links on the page. You need to format the annotations in the following format:

    /path/to/video_1 label_1
/path/to/video_2 label_2

We provide a script for this (scripts/make_annot_file.py) to construct the annotation file from the video folder structure, where each video is placed in a folder with the same name as the label.

Once complete, make a directory in the main folder called data and move the extracted data into it. the project directory structure should look like this:

.
├── rlt
│   ├── src
│   ├── data
│   │   ├── kinetics400
│   │   │   ├── train_labels.txt
│   │   │   ├── val_labels.txt
│   └── ...

To check that all the data is loading properly, we advise you run the benchmark_dataloader.py script:

python scripts/benchmark_dataloader.py

If this can iterate through the whole dataset, everything is set up properly!

Something-Something V2

  1. Please download the dataset and annotations from dataset provider.

  2. Download the frame list from the following links: (train, val).

  3. Run scripts/reformat.py to create the metadata file.

  4. How to clean labels (from Sihan) (TODO)

Setup Datasets Path

For ease of use, we recommend setting the following environment variables to avoid having to specify the data path in the config file.

Kinetics400

export KINETICS_TRAIN_METADATA=/yourpath/kinetics400_train.txt
export KINETICS_VAL_METADATA=/yourpath/kinetics400_val.txt

SSV2

export SSV2_TRAIN_ANNO=/yourpath/train.csv
export SSV2_VAL_ANNO=/yourpath/train.csv