People shamed and ordered to leave shops after being misidentified then ‘given no help’ to investigate verdicts

  • ryven@lemmy.dbzer0.com
    link
    fedilink
    English
    arrow-up
    36
    ·
    edit-2
    8 hours ago

    I suspect the claimed “99.98% accuracy” is counting out of all faces scanned, which is a bullshit way to make the tech look good. Most faces are not marked as shoplifters in the database. A system that literally does nothing would probably still have greater than 99% accuracy.

    What we really want to know is what percentage of reported matches are accurate, and I bet it isn’t anywhere near 99%.

    • Treczoks@lemmy.world
      link
      fedilink
      English
      arrow-up
      7
      ·
      6 hours ago

      I had to try to educate sales people what such numbers actually mean.

      With fingerprint readers, there are false positives (your finger is accepted, although it should not), and false negatives (your finger gets rejected although it should accept). The chances for both look small, but if you have 700+ people in the system, the chance of a random person to be accepted as one of the 700 is about bigger than 50%. And there was a big chance for any valid user to be logged in as someone else.

      • RunawayFixer@lemmy.world
        link
        fedilink
        English
        arrow-up
        3
        ·
        edit-2
        2 hours ago

        Pretty much this. A 0.02% error margin when there are tens of thousands of visitors per year, means it’s almost guaranteed to have errors.

        99.9% ^700 = 49.6% chance of no errors occurring.
        99.98% ^3466 = 50% chance of no errors occurring.
        99.98% ^23000 = 1% chance of no errors occurring.