This repository contains the code for the paper:

Multi-query Video Retreival

@misc{wang2022multiquery,
      title={Multi-query Video Retrieval}, 
      author={Zeyu Wang and Yu Wu and Karthik Narasimhan and Olga Russakovsky},
      year={2022},
      eprint={2201.03639},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

1. Download raw videos for MSR-VTT, MSVD and VATEX, and put them into data/{dataset}/raw_videos folder.
2. Run the script data/extract_frames.sh to extract frames from raw videos.

The resulting data folder structures like this:

├── data
    ├── msrvtt
        ├── msrvtt_train.json
        ├── msrvtt_test.json
        ├── msrvtt_test_varying_query_sample_1-20.json
        ├── raw_videos
            ├── video0.mp4
            ├── ...
        ├── extracted_frames
            ├── video0.mp4
                ├── 0.jpg
                ├── ...
            ├── ...
    ├── msvd
        ├── ...
    ├── vatex
        ├── ...

For Frozen model, download the pretrained checkpoint provided by the original authors here, and put into record/pretrained folder.

Run command: python train.py -c configs/{config_path}

Run command: python evaluate.py -c configs/{config_path}

The structure of this repository is based on https://github.com/victoresque/pytorch-template. Some of the code are adpated from https://github.com/m-bain/frozen-in-time and https://github.com/ArrowLuo/CLIP4Clip.