Skip to content

Latest commit

 

History

History
83 lines (57 loc) · 3.52 KB

DATA.md

File metadata and controls

83 lines (57 loc) · 3.52 KB

Dataset Preparation

Kinetics-400

Since some urls of Kinetics-400 dataset are expired, we cannot collect full dataset.
Alternatively, we downloaded the dataset via a link from the facebookresearch/video-nonlocal-net#67.
However, the link is also expired and thus we share our dataset link.

[Kinetics-400 dataset link]

  • training set : 234,619 videos
  • validation set : 19,716 videos

Important : This dataset is smaller than the original Kinetics-400 dataset.

After all the videos were downloaded, prepare the csv files for training and validation as train.csv, val.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

For an example sample, we also share our csv.files including our absolute path. So you have to replace our absolute path to your path.

For dataset specific issues, please reach out to the dataset provider.

Something-Something-V1 (SSv1)

Step 1. Download the dataset and annotations from dataset provider.
*Note that unlike SSv2 (video files), SSv1 is comprised of extracted RGB frames.

cd path/to/the/data/ssv1
cat 20bn-something-something-v1-?? | tar zx

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

Please put all annotation json files and the frame lists in the same folder, and set DATA.PATH_TO_DATA_DIR to the path. Set DATA.PATH_PREFIX to be the path to the folder containing extracted frames.

cp *.csv 20bn-something-something-v1/

Something-Something-V2 (SSv2)

Step 1. Download the dataset and annotations from dataset provider.

cd path/to/the/data/ssv2
cat 20bn-something-something-v2-?? | tar zx

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

Step 3. Extract the frames at 30 FPS by using video_extractor_sthv2.py.
*Note that check the number of the extracted frames : 25,209,271
*If you get the fewer number of frames than mine, remove the folder and try it again.

# Extract videos to frames
python vov3d/datasets/video_extractor_sthv2.py \
--video_dir path/to/the/data/ssv2/20bn-something-something-v2 \
--frame_dir path/to/the/data/ssv2/20bn-something-something-v2-SF-frame

# Check the number of extracted frames : 25209271
cd path/to/the/data/ssv2/20bn-something-something-v2-SF-frames
find . -type f | wc -l

Please put all annotation json files and the frame lists in the same folder, and set DATA.PATH_TO_DATA_DIR to the path. Set DATA.PATH_PREFIX to be the path to the folder containing extracted frames.

ETC

For AVA and Charades, please follow PySlowFast's instruction.