Skip to content

DATASETS.md

Latest commit

 

History

History
64 lines (33 loc) · 2.71 KB

DATASETS.md

File metadata and controls

64 lines (33 loc) · 2.71 KB
Raw

Datasets

We considered 12 video datasets in total for our experiments.

🔗 Downloading Metadata

To simplify processing, we aligned the metadata across all datasets.
You can download it from the following link and place it in the ./metadata directory:
🔗 GDrive link.

Note: You may need to adjust the full_path entries depending on where and how you store your videos.

🧪 Test Splits Used

Following prior work, we use the following splits for evaluation:

  • test splits: msrvtt, didemo, lsmdc, S-MiT

  • val splits: kinetics_400, kinetics_600, kinetics_700, MiT, activity_net, youcook, ssv2

🎞️ Downloading Videos

  • MSRVTT, LSMDC, and DiDeMo: Follow the CLIP4CLIP guidelines to download videos of these datasets. We store videos in the following folders:

    • data/msrvtt/videos/all/

    • data/lsmdc/avi/ (with subfolders)

    • data/didemo/videos/

  • Kinetics 400, 600, 700: Follow the Kinetics dataset instructions to download. We store videos in the following folders:

    • data/kinetics_400/{train,val}_data_resized/ (with class subfolders)

    • data/kinetics_600/videos/{train,val,test}/ (with class subfolders)

    • data/kinetics_700/kinetics-dataset/k700-2020/{train,val,test}/ (with class subfolders)

  • MiT and S-MiT: Follow the Moments in Time website for download instructions. We store videos in the following folders:

    • data/MiT/Moments_in_Time_Raw/{training,validation}/ (with class subfolders)

    • data/S-MiT/videos/ (with class subfolders)

  • ActivityNet: Follow the official ActivityNet download instructions. Access the full dataset via Google Drive or Baidu after submitting a request. We used only the downsampled version Anet_videos_15fps_short256:

    • data/activity_net/Anet_videos_15fps_short256/

  • YouCook2: Follow the YouCook2 guidelines to download. We store videos in the following folders:

    • data/youcook/{training,validation,testing}/

  • UCF101: Download from the official UCF101 website. Store the videos in:

    • data/UCF101/original/videos/

  • Something-Something v2: Follow the official instructions. We converted all videos to .mp4 and stored them in:

    • data/SomethingSomething_v2/ssv2/train_data_mp4_correct/

    • data/SomethingSomething_v2/ssv2/val_data_mp4_correct/