3 ms·

TikTok dances trained an AI to see

• ryzvonusef 4y ago
  > The quest for computer vision requires lots of data — including real world images. But that can be hard to find, which has led researchers to look in some pretty creative places. > The above video shows how researchers used Tik Tok dances and the Mannequin Challenge to train AI. The quest is for “ground truth” — real world examples that can be used to train or grade an AI on its guesses. Tik Tok datasets provide this by showing lots of movement, clothing types, backgrounds, and people. That diversity is key to train a model that can handle the randomness of the real world. > The same thing happens with the Mannequin Challenge — all those people pretending to stand still gave researchers — and their models — more real world data to train with than they ever could have hoped for. > Further Reading: > Tik Tok aided depth: https://www.yasamin.page/hdnet_tiktok > Mannequin Challenge: https://google.github.io/mannequinchallenge/www/index.html > Geofill and Reference-Based Inpainting: https://paperswithcode.com/paper/geofill-reference-based-image-inpainting-of > Virtual Correspondence: https://virtual-correspondence.github.io/ > Densepose: http://densepose.org/