This repository contains the code for our ICCV 2021 paper:

ACAV100M: Automatic Curation of Large-Scale Datasets for Audio-Visual Video Representation Learning
Sangho Lee*, Jiwan Chung*, Youngjae Yu, Gunhee Kim, Thomas Breuel, Gal Chechik, Yale Song (*: equal contribution)
[paper]

@inproceedings{lee2021acav100m,
    title="{ACAV100M: Automatic Curation of Large-Scale Datasets for Audio-Visual Video Representation Learning}",
    author={Sangho Lee and Jiwan Chung and Youngjae Yu and Gunhee Kim and Thomas Breuel and Gal Chechik and Yale Song},
    booktitle={ICCV},
    year=2021
}

On our official website (https://acav100m.github.io/), we provide more information on the dataset including following features:

• Dataset download links
• Sample video clip explorer

• Python >= 3.8.5
• FFMpeg 4.3.1

1. Install PyTorch 1.6.0, torchvision 0.7.0 and torchaudio 0.6.0 for your environment. Follow the instructions in HERE.

2. Install the other required packages.

pip install -r requirements.txt
python -m nltk.downloader 'punkt'
pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/<cuda version>/torch1.6/index.html
pip install git+https://github.com/jiwanchung/slowfast
pip install torch-scatter==2.0.5 -f https://pytorch-geometric.com/whl/torch-1.6.0+<cuda version>.html

e.g. Replace <cuda version> with cu102 for CUDA 10.2.

1. Create the data directory

mkdir data

2. Prepare the input file.

data/metadata.tsv should be structured as follows. We provide an example input file in examples/metadata.tsv

YOUTUBE_ID\t{"LatestDAFeature": {"Title": TITLE, "Description": DESCRIPTION, "YouTubeCategory": YOUTUBE_CATEGORY, "VideoLength": VIDEO_LENGTH}, "MediaVersionList": [{"Duration": DURATION}]}

bash ./run.sh

To enable GPU computation, modify the CUDA_VISIBLE_DEVICES environment variable accordingly. For example, run the above command as export CUDA_VISIBLE_DEVICES=2,3; bash ./run.sh.

1. Filter the videos with metadata.

bash ./metadata_filtering/code/run.sh

The above command will build the data/filtered.tsv file.

2. Download the actual video files from youtube.

bash ./video_download/code/run.sh

Although we provide a simple download script, we recommend more scalable solutions for downloading large-scale data.

The above command will download the files to data/videos/raw directory.

3. Segment the videos into 10-second clips.

bash ./clip_segmentation/code/run.sh

The above command will save the segmented clips to data/videos directory.

4. Extract features from the clips.

bash ./feature_extraction/code/run.sh

The above command will save the extracted features to data/features directory.

This step requires GPU for faster computation.

5. Perform clustering with the extracted features.

bash ./clustering/code/run.sh

The above command will save the extracted features to data/clusters directory.

This step requires GPU for faster computation.

6. Select subset with high audio-visual correspondence using the clustering results.

bash ./subset_selection/code/run.sh

The above command will save the selected clip indices to data/datasets directory.

This step requires GPU for faster computation.

The final output should be saved in the data/output.csv file.

output.csv is structured as follows. We provide an example output file at examples/output.csv.

# SHARD_NAME,FILENAME,YOUTUBE_ID,SEGMENT
shard-000009,qpxektwhzra_292.mp4,qpxektwhzra,"[292.3329999997, 302.3329999997]"

Instructions on downstream evaluation are provided in Evaluation.

Instructions on correspondence retrieval experiments are provided in Correspondence Retrieval.