Detroit woman falsely arrested while pregnant due to facial recognition tech | Hacker News Reader

Detroit woman falsely arrested while pregnant due to facial recognition tech

nbcnews.com

44 pointsnotRobot21 comments

Let me guess, she's black...

Aaaand sure enough, she's black. Facial recognition is kinda crap for anyone who isn't white (and probably male). A problem from training sets and building the models, no doubt.

Blindly trusting facial recognition is a really, really bad idea for the near and mid-term future.

It's also a problem of lower signal to noise ratio: darker faces reflect less light and it's harder to extract data on key facial features like eyebrows. No amount of training or fine tuning is going to make an f-22 easier to detect than an f-16 for the same reason.

> darker faces reflect less light and it's harder to extract data on key facial features like eyebrows.

Pretty much this. It gets even worse under non-ideal lighting conditions.

I mean, me? I pretty much glow in the dark. You’d be hard-pressed to find a whiter guy short of an albino. But I have friends who can practically disappear in dusk/twilight conditions, and the only way you’ll ever see them in the dark is if they look directly at you or smile.

And I have _great_ night vision. Facial recognition cameras? Yeah, not so much.

I think that saying a black face has less signal than a white face is, while true, not quite relevant. Because it's more about the difference in luminosity across features and not the absolute luminosity.

The best solution is to get cameras which can see and record both visible and non-visible light spectrums, since they will provide more uniform contrasts for edge detection across skin tones.

But until then, we need to stop blindly trusting the software.

The luminosity is the signal:

> difference in luminosity across features and not the absolute luminosity

This is exactly what signal to noise ratio refers to in my previous comment.

The best solution is what the iPhone does: it is a 3d sensor that produces a depth map (think each pixel represents a distance to the object). But that's not really viable when trying to use normal security cameras.

In our visible spectrum. Maybe (I don't know) black faces have more contrast in some other "non-visible" wavelength, and it's just a matter of using the right camera.

I did imaging on skin (for medical appliances, not for surveillance that I heavily oppose). While the tech has evolved a lot in just a few years, a tiny bit of additional light exposure can make a huge difference. I doubt surveillance cameras do quality images with a high dynamic range of light levels.

Seems simple to me - the software should just decline to match if there isn't enough data.

Improve the training data. There's a large set of photographs with front and profile views. Black men make up a large fraction of this set. Might still leave a gap in recognizing women so society should try to add more women to that dataset.

Hum... Wait, the police made the software compare some (probably low resolution) video against the picture of (probably literally) everybody and came-up with a name? And then simply believed on that name run after the person and arrested her?

That's ridiculous.

The victim identified her from a photo lineup, but the police used a picture of her from an arrest 8 years earlier rather than her latest driver's license photo that they could also access. I think the problem here is that if you have enough faces to compare, you're going to find some similar looking ones and a victim being shown a lineup will probably assume one of those faces must belong to the perp. I does seem ridiculous if that's all they had for the arrest.

Everything else aside, why is pregnancy an important part of the article's title? This is terrifying regardless of whether you're pregnant or not.

Because the charge was for carjacking and she was eight months pregnant at the time. You're right that the facial recognition aspect was bad enough, but there was a very obvious way they could have verified their arrest with the victim "was she obviously heavily pregnant"?

To add to the reader's sense of incredulity that this person was the perpetrator of the crime.

Not even sure what crime they are looking at here.

The police said they suspected her to be a woman who returned a stolen phone.

That does not sound like something a thief would do after carjacking someone.

The person they were looking for was probably also not a criminal.

The carjacker was not visibly pregnant, which makes this woman's arrest even more ridiculous

So she went to a lineup using pictures (in this case an 8 year old photo) and the victim chose her as the suspect? Isn't a lineup supposed to have the actual suspect and others lined up? How can they use old photos?

This is a miscarriage of justice and I hope the police get raked over the coals for using BS facial recognition in a live case and ignoring circumstantial evidence.

Dupe from yesterday: https://news.ycombinator.com/item?id=37025650

It's only a matter of time until someone is killed becuase of an incorrect identification from a facial recognition or AI/ML system.

True, but people have already been killed plenty of times through history from incorrect identification by human facial recognition.

So there are other systems and checks-and-balances that need improvement too.

The legacy criminal system and all the laws around background checks are not ready for the false positives of AI.

Either legacy criminal laws need to be changed or tech companies need to be made liable for false positives.

Reply on news.ycombinator.com