Short Video Dataset



About

SVD is a large-scale short video dataset, which contains over 500,000 short videos collected from http://www.douyin.com and over 30,000 labeled pairs of near-duplicate videos.

News

  • Our paper "SVD: A Large-Scale Short Video Dataset for Near-Duplicate Video Retrieval" was accepted by ICCV-2019.

Publications

  • SVD: A Large-Scale Short Video Dataset for Near Duplicate Video Retrieval. [pdf]
          Qing-Yuan Jiang, Yi He, Gen Li, Jian Lin, Lei Li and Wu-Jun Li.
          Proceedings of International Conference on Computer Vision (ICCV), 2019.
  • Citation

    Please cite the following papers if you use this dataset.
     @inproceedings{JiangHLLLL2019,
       title={{SVD}: A Large-Scale Short Video Dataset for Near-Duplicate Video Retrieval},
       author={Qing-Yuan Jiang, Yi He, Gen Li, Jian Lin, Lei Li and Wu-Jun Li},
       booktitle={Proceedings of International Conference on Computer Vision},
       year={2019}
     }

    Download

    Download Instruction

    • The SVD dataset is available for non-commercial research purposes only.
    • Please download the agreement and read it carefully.
    • Please ask your supervisor/advisor to sign the agreement appropriately and then send the scanned version (example) to Jian Lin [linj(at)lamda.nju.edu.cn] and cc to Yi He [heyi_main(at)126.com].
    • After verifying your request, we will contact you with the password to unzip the metadata.zip and videos-urls.zip.

    Labeled Pairs and Video URLS

    Download Demo

    • We provide the download demo to download the videos based on aformentioned urls. Please refer to github repo: [github]

    Preprocessing

    Transformed Demos

    We define four transformations for the SVD dataset. Here we present an example:

    Original Video Video Speeding Video Cropping Black Border Insertion Video Rotation

    We provide the source codes to perform these transformations, please refer to the repo: [github].

    Implementations of the baselines

    We provide the implementations of the baselines for the SVD dataset, please refer to the repo: [github].

    LeaderBoard for Near-Duplicate Video Retrieval Task

    Real-value based Methods
    Method Code MAP@SVD
    N/A Cropping Black Border Rotation Speeding
    1
    N/A
    DML
    G.Kordopatis-Zilos, et al.
    78.47 54.07 68.17 15.59 76.70
    2
    N/A
    CNNL
    G. Kordopatis-Zilos, et al.
    55.55 15.61 18.63 0.15 51.80

    Contact

    If you have any questions about this dataset, please contact: