• AgreeableLandscapeOP
    link
    fedilink
    arrow-up
    2
    ·
    3 years ago

    In theory they could have used some public domain datasets or even parts of Wikimedia Commons.

    • Arthur BesseMA
      link
      fedilink
      arrow-up
      3
      ·
      3 years ago

      It appears that the captioning model on that website was trained on the MSCOCO dataset which was sourced from from Google and Bing image search, and also from Flickr.