ECLIPSE: Efficient Long-range Video Retrieval using Sight and Sound

This is the PyTorch implementation of our paper:
ECLIPSE: Efficient Long-range Video Retrieval using Sight and Sound
Yan-Bo Lin, Jie Lei, Mohit Bansal, and Gedas Bertasius
In European Conference on Computer Vision, 2022.

paper

📝 Preparation

1. pip3 install requirements.txt
2. Dataset: ActivityNet, QVHighlights, YouCook2, DiDeMo and Charades.
3. extract video frames in 3 fps.
4. extract audio features.
5. To load pretrained CLIP weight

The download links are from official CLIP4Clip Download CLIP (ViT-B/32) weight,

wget -P ./modules https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt

or, download CLIP (ViT-B/16) weight,

wget -P ./modules https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt

💿 Extract images and audio features.

ActivityNet/
├── raw_frames/
│       └── VIDEO_NAME/
│           ├── 0001.jpg
│           ├── ...
│           └── 00...jpg
│
└── VGGSound_Audio_features_10s_aligned/
        └── VIDEO_NAME/
            ├── 0000.pt
            ├── ...
            └── 00...pt

💿 Extracted audio features.

VGGSound features on ActivityNet Captions: Google Drive

📚 Train and evaluate

ActivityNet Captions: bash run_act.sh
DiDemo: bash run_didemo.sh
Charades: bash run_cha.sh
QVHighlight:bash run_qvh.sh
YouCook2: bash run_yc2.sh

🎓 Cite

If you use this code in your research, please cite:

@InProceedings{ECLIPSE_ECCV22,
author = {Yan-Bo Lin and Jie Lei and Mohit Bansal and Gedas Bertasius},
title = {ECLIPSE: Efficient Long-range Video Retrieval using Sight and Sound},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
month = {October},
year = {2022}
}

👍 Acknowledgments

Our code is based on CLIP4Clip and VGGSound

✏ Future works

• Preprocessed video frames and audio features

License

This project is licensed under MIT License, as found in the LICENSE file.