And the dataset is prbably racist, although in the reported case, it sounds like good old unreliable cross-race recognition by humans, with the evil eye pinging because it spotted someone and the store staff then telling the wrong person to naff off. It seems like a process or training failure if they don’t ask the evil eye to confirm they’ve got the person it flagged before upsetting them.
Comment on Facial recognition error: Customer misidentified by Sainsbury's
Zombie@feddit.uk 1 day ago
Both Facewatch and Sainsbury’s point to the software’s “99.98% accuracy” – but Rajah suspects the margin of error is higher and has questions about the dataset behind this claim, and if it is representative of a range of body types and skin colours.
99.98% looks good to a layman, but that number is meaningless in reality.
Is that 0.02% error false positives or false negatives, or both?
Also, 0.02% means 2 in every 10,000. I don’t think it takes long for 10,000 people to go through the doors of Sainsburys every day, considering the UK population is about 65 million and they’re a nationwide company. Once this is rolled out nationwide they’re going to have constant false flags.
Scumbag oppressive tactics by a scumbag company.
mjr@infosec.pub 1 day ago
halcyoncmdr@piefed.social 1 day ago
Yeah, 0.02% of 65 Million is 1.3 Million possible errors.
And that’s just based on the raw population, that accuracy rating could be based on raw number of scans instead. A quick search shows Sainsbury’s serves 16 million customers a week. That’s 320,000 errors every week if the error rate is just raw scans as opposed to unique scans.
Nollij@sopuli.xyz 19 hours ago
Your math is off, by two decimal places.
Zombie@feddit.uk 15 hours ago
Indeed, it’s still a ridiculous amount of errors though.
65,000,000 x (0.02%) = 13,000 possible errors
16,000,000 x (0.02%) = 3,200 errors every week