Detroit police chief admits facial recognition is wrong “96% of the time” | Hacker News Reader

Detroit police chief admits facial recognition is wrong “96% of the time”

arstechnica.com

55 pointsfurcyd8 comments

Unscientific. The best facial recognition is accurate 98%-99.9% of the time. There are minor known problems with blacks (~2% worse), bad lighting, bad cameras, but big problems with small databases to compare against. There are yearly vendor tests and challenges. E.g https://en.wikipedia.org/wiki/Face_Recognition_Grand_Challen... Latest winners are all in the 99.9% ballpark, not 4%.

If the Detroit police has no faces to compare against and the image quality is crap, it's certain that their success rate cannot be much better than 10%.

4% success rate is fantastic for one tool bring the perpetrator to the attention of investigators.

The problem is in abusing the tool as in the watch theft case.

More like 96% distraction.

Depends on how many results it brings you -- if it declares 10 candidate matches, and 1 is correct, then its still 90% wrong -- but still very useful.

And it ruins the lives of the 9 incorrect matches, because of the way the legal system works in America. I'd say that that's 9 lives too many.

Is this life ruining thing a thing, or is it a generalizing of that episode that hit the news a week ago, or is it an instantiation of the current zeitgeist according to which there is no ill that can’t be associated with US law enforcement

I don't know why but for some reason US law enforcement always manages to find a way to maximize harm.

When you give them less lethal weapons what happens is that they just use them in every possible situation and often they increase lethality by aiming at vulnerable body parts or shoot them from a very short distance. What they really should do is show some restraint and not let themselves get provoked easily.

Facial recognition is basically ending in the same situation. 96% failure rate means you have a lead but not evidence and only get to interview people for more clues. Instead they just arrest the leads because the wanted criminal could attack during the interview or flee afterwards.

It's simply not realistic to expect highly accurate facial recognition from public surveillance cameras.

Resolution is often lacking, angles are distorted, illumination is poor, people are moving, features are hidden or blurred, people wear hats, etc., etc..

Mug shots, maybe. Frame grabs from public surveillance, not very likely.

Reply on news.ycombinator.com