Permanently Deleted

  • Budwig_v_1337hoven [he/him]
    ·
    3 years ago

    not for the actual recognition I imagine, but for training a model you'd want your training data to be somewhat consistent (faces in the middle, no rotation, neutral background, similar lighting etc)

    • QuillcrestFalconer [he/him]
      ·
      3 years ago

      Not really, if you want a robust recognition system. In fact it's often normal in image sets to create part of your training data by transforming pictures of your training set, for example by rotating, translating, scaling etc

      • Budwig_v_1337hoven [he/him]
        ·
        3 years ago

        really depends on what exactly your doing, I'd say. Sure, variety is generally good but only in so far as it's actually required in the particular use case.
        All I'm saying is this doesn't seem like an entirely implausible explanation. You can even see the old image under the rotated one - not exactly looking like a typical AI generated media glitch, from what I've seen so far.

    • Ideology [she/her]
      ·
      3 years ago

      So, my question, then, would be to ask why they're using their target pop in a training set when the system is live.

      • Budwig_v_1337hoven [he/him]
        ·
        3 years ago

        If I'd have to guess, I'd say refinement. Further training in an effort to get accuracy up. As good as these systems are, they're never fully accurate, never truly 'finished'

        All that being said, this is just me taking some stabs in the dark.