cleanvid-15m_map / README.md
chavinlo's picture
Update README.md
59fbdfc
|
raw
history blame
4.9 kB
metadata
license: cc-by-4.0
dataset_info:
  features:
    - name: id
      dtype: int64
    - name: description
      dtype: string
    - name: duration
      dtype: float64
    - name: aspectratio
      dtype: string
    - name: videourl
      dtype: string
    - name: author
      dtype: string
    - name: categories
      dtype: string
    - name: framerate
      dtype: float64
    - name: r18
      dtype: int64
  splits:
    - name: train
      num_bytes: 16755833083
      num_examples: 14394510
  download_size: 5410262648
  dataset_size: 16755833083
task_categories:
  - text-to-video
  - video-classification
language:
  - en
tags:
  - captions
  - metadata
pretty_name: CleanVid Map (15M)
size_categories:
  - 10M<n<100M

CleanVid Map (15M) πŸŽ₯

TempoFunk Video Generation Project

CleanVid-15M is a large-scale dataset of videos with multiple metadata entries such as:

  • Textual Descriptions πŸ“ƒ
  • Recording Equipment πŸ“Ή
  • Categories πŸ” 
  • Framerate 🎞️
  • Aspect Ratio πŸ“Ί

CleanVid aim is to improve the quality of WebVid-10M dataset by adding more data and cleaning the dataset by dewatermarking the videos in it.

This dataset includes only the map with the urls and metadata, with 3,694,510 more entries than the original WebVid-10M dataset. Note that the videos are low-resolution, ranging from 240p to 480p. But this shouldn't be a problem as resolution scaling is difficult in Text-To-Video models. More Datasets to come for high-res use cases.

CleanVid is the foundation dataset for the TempoFunk Video Generation project.

Built from a crawl of Shutterstock from June 25, 2023.

Format πŸ“Š

  • id: Integer (int64) - Shutterstock video ID
  • description: String - Description of the video
  • duration: Float(64) - Duration of the video in seconds
  • aspectratio: String - Aspect Ratio of the video separated by colons (":")
  • videourl: String - Video URL for the video in the entry, MP4 format. WEBM format is also available most of the times (by changing the extension at the end of the URL.).
  • author: String - JSON-String containing information of the author such as Recording Equipment, Style, Nationality and others.
  • categories: String - JSON-String containing the categories of the videos. (Values from shutterstock, not by us.)
  • framerate: Float(64) - Framerate of the video
  • r18: Bit (int64) - Wether the video is marked as mature content. 0 = Safe For Work; 1 = Mature Content

Code πŸ‘©β€πŸ’»

If you want to re-create this dataset on your own, code is available here:

https://github.com/chavinlo/tempofunk-scrapper/tree/refractor1/sites/shutterstock

Due to rate-limitations, you might need to obtain a proxy. Functionality for proxies is included in the repository.

Sample πŸ§ͺ

{
  "id": 1056934082,
  "description": "Rio, Brazil - February 24, 2020: parade of the samba school Mangueira, at the Marques de Sapucai Sambodromo",
  "duration": 9.76,
  "aspectratio": "16:9",
  "videourl": "https://www.shutterstock.com/shutterstock/videos/1056934082/preview/stock-footage-rio-brazil-february-parade-of-the-samba-school-mangueira-at-the-marques-de-sapucai.mp4",
  "author": {
    "accountsId": 101974372,
    "contributorId": 62154,
    "bio": "Sempre produzindo mais",
    "location": "br",
    "website": "www.dcpress.com.br",
    "contributorTypeList": [
      "photographer"
    ],
    "equipmentList": [
      "300mm f2.8",
      "24-70mm",
      "70-200mm",
      "Nikon D7500 ",
      "Nikon Df",
      "Flashs Godox"
    ],
    "styleList": [
      "editorial",
      "food",
      "landscape"
    ],
    "subjectMatterList": [
      "photographer",
      "people",
      "nature",
      "healthcare",
      "food_and_drink"
    ],
    "facebookUsername": "celso.pupo",
    "googlePlusUsername": "celsopupo",
    "twitterUsername": "celsopupo",
    "storageKey": "/contributors/62154/avatars/thumb.jpg",
    "cdnThumbPath": "/contributors/62154/avatars/thumb.jpg",
    "displayName": "Celso Pupo",
    "vanityUrlUsername": "rodrigues",
    "portfolioUrlSuffix": "rodrigues",
    "portfolioUrl": "https://www.shutterstock.com/g/rodrigues",
    "instagramUsername": "celsopupo",
    "hasPublicSets": true,
    "instagramUrl": "https://www.instagram.com/celsopupo",
    "facebookUrl": "https://www.facebook.com/celso.pupo",
    "twitterUrl": "https://twitter.com/celsopupo"
  },
  "categories": [
    "People"
  ],
  "framerate": 29.97,
  "r18": 0
}

Credits πŸ‘₯

Main

  • Lopho - Part of TempoFunk Video Generation
  • Chavinlo - Part of TempoFunk Video Generation & CleanVid Crawling, Scraping and Formatting
@InProceedings{Bain21,
  author       = "Max Bain and Arsha Nagrani and G{\"u}l Varol and Andrew Zisserman",
  title        = "Frozen in Time: A Joint Video and Image Encoder for End-to-End Retrieval",
  booktitle    = "IEEE International Conference on Computer Vision",
  year         = "2021",
}

Extra

  • Salt - Base Threading Code (2022)