3 ms·

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 t

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.

m463 3y ago
Seems simple to me - the software should just decline to match if there isn't enough data.
Am4TIfIsER0ppos 3y ago
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.
Manuel_D 3y ago
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.
- falcolas 3y ago
  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.
  - raxxorraxor 3y ago
    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.
  - Manuel_D 3y ago
    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.
    - bestouff 3y ago
      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.
- rekabis 3y ago
  > 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.

m463 3y ago

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

Am4TIfIsER0ppos 3y ago

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.

Manuel_D 3y ago

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.

falcolas 3y ago
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.
- raxxorraxor 3y ago
  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.
- Manuel_D 3y ago
  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.
  - bestouff 3y ago
    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.
rekabis 3y ago
> 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.