4 ms·

In my naive mind, I can't help but think it might simply be a case of the contrast ratio being greater in lighter skinned people - therefore making the hit rati

by sundvor 10y ago

In my naive mind, I can't help but think it might simply be a case of the contrast ratio being greater in lighter skinned people - therefore making the hit ratios higher. Greater contrast ratios will make facial features stand out more.

handedness 10y ago
Exactly my thinking. I worked with a company that developed on a medical device that took subcutaneous readings extracutaneously, and despite being developed almost entirely by Asian engineers, it got the best data quality from fair-skinned Caucasians, and the worst from dark-skinned Africans, as it totally relied on the transmission of light through skin tissue. Claiming racism without understanding the problem is both unfair to the people involved, and an unwise undermining of one's own credibility for the times when people really are acting in bad faith. That may be in this case, but that may not be, and I don't like guilty-until-providen-innocent no matter whom it's leveled against (even a Federal agency of which I'm not the world's biggest fan, and a database the likes of which make me think we need a constitutional amendment to protect the privacy of the citizens).
- mambodog 10y ago
  Even if it was no-one's intent that the system is racially biased, allowing such a system to continue operate after learning of such a bias is a racist act.
  - lgessler 10y ago
    I'm inclined to agree with you, but what would you say if it turned out that after much effort the engineering team was unable to make performance equal across all skin tones?
    - jacquesm 10y ago
      You mean that it would work better with darker skin?
      - handedness 10y ago
        What if the technical challenges of identifying a person's face mean it will always be challenging to identify someone when the light and/or contrast levels are lower? What then? Manually cripple the system and misidentify white suspects, too? This post's comments are a great example of why technologists are fairly terrible at engineering social solutions, much the same way sociologists tend to be terrible at engineering technical solutions.
  - handedness 10y ago
    Ignoring known technical deficiencies in the system and assigning the same confidence levels in the output regardless of race could work out to be pretty racist, but what are they supposed to do, manually cripple the accuracy of the system down to the lowest common denominator? In the medical device I mentioned above, it was medically useful for caucasians and lighter-skinned Hispanics, so-so with Asians, and almost useless for Africans. So, what–don't go to market until the technology has parity across all skin tones? Not everything is about intent.
    - bilbo0s 10y ago
      I hate to be a pedant here... but I actually wrote a pretty successful 3d radiology/cardiology package. So I understand medical devices. I just wanted to hop in here and say that there is NO WAY the FDA would give you clearance to use the device you describe on African Americans under the circumstances that your posts have postulated. It seems a no brainer that the facial recognition software should not be used on African Americans either. At least, if it was MEDICAL software... it would DEFINITELY NOT be cleared with that kind of a failure rate. And any provider or delivery network using the facial recognition software on African Americans would be on his or her way to prison. But the regulations on legal software may be more lackadaisical than the regulations on medical software. Getting things exactly right may not be as important. And the penalties for misuse of medical devices are, without question, far more draconian.
      - handedness 10y ago
        > I just wanted to hop in here and say that there is NO WAY the FDA would give you clearance to use the device you describe on African Americans under the circumstances that your posts have postulated. The FDA has approved drugs that are specifically intended to target medical issues in specific ethnicities, so I'm not sure about your assessment. That said, this product's market was outside the scope of the FDA, so it's irrelevant. > It seems a no brainer that the facial recognition software should not be used on African Americans either. Any data source which produces data that's better than a random guess should be considered when trying to solve a problem. That it offends some people's social sensibilities is a separate issue. Look, I don't like the mere existence of the FBI's system, and I don't think the public's continued lack of concern for placing constitutional restraints on systems like it is a wise course forward, but to say it's a "no-brainer" when it might produce valuable leads is to focus on one aspect of the problem without considering the rest of it. Just as it would also be a mistake to consider it more accurate than it is when trying to identify dark-skinned suspects. Witness accuracy is not a new problem. It's also worth noting that the human eye is just as bad at facial recognition when the light and/or contrast is lower.
chillacy 10y ago
Depends on the camera settings too. Older movies with mostly white casts messed this up a lot, by not properly calibrating for the darker skinned characters. Conversely, movies with all black casts get this right: they tune their camera/color settings to make the faces more distinguishable. I ran into this a few weeks ago where I took a picture of a couple friends. The camera decided to pick the friend with dark skin, resulting in him looking great, and my white friend (who isn't even super pale or anything) getting washed out.
- sundvor 10y ago
  Yeah, I shoot semi-manual on my old D700 SLR. If you want high contrast ratios in dark areas of an image, everything else will be blown out - and you'll have to use a longer shutter speed - or suffer iso degradation or blur.
  - chillacy 10y ago
    > If you want high contrast ratios in dark areas of an image, everything else will be blown out Conversely could you say that in order to capture subjects with lighter skin, you need to increase lighting in the background or else the background will clip? Here's an article about the cinematography and technology choices: https://www.washingtonpost.com/entertainment/movies/12-years-a-slave-mother-of-george-and-the-aesthetic-politics-of-filming-black-skin/2013/10/17/282af868-35cd-11e3-80c6-7e6dd8d22d8f_story.html https://www.washingtonpost.com/entertainment/movies/12-years... > he was able to photograph black actors of dramatically different skin tones in a nighttime interior scene using just everyday house lamps, thanks to a sophisticated digital camera. “I just changed the wattage of the bulb, used a dimmer, and I didn’t have to use any film lights. That kind of blew me away,” Patterson says. “The camera was able to hold both of them during the scene without any issues.” It seems like there might have been tech limitations initially, but with today's technology you can light your backgrounds and expose for either dark or light subjects well. The big challenge is having a photo with 2 subjects in the same exposure who you want to both look good. Edit: by the way I too have an old Nikon camera. I usually shoot in Aperture priority with some intentional underexposure, and then edit it up in post :)
ma2rten 10y ago
If it's a machine learning system (which virtually all face recognition systems are as far as I know) it depends more than anything on the training data. It's just the same as with people: https://en.wikipedia.org/wiki/Cross-race_effect https://en.wikipedia.org/wiki/Cross-race_effect

by sundvor 10y ago

handedness 10y ago

Exactly my thinking. I worked with a company that developed on a medical device that took subcutaneous readings extracutaneously, and despite being developed almost entirely by Asian engineers, it got the best data quality from fair-skinned Caucasians, and the worst from dark-skinned Africans, as it totally relied on the transmission of light through skin tissue. Claiming racism without understanding the problem is both unfair to the people involved, and an unwise undermining of one's own credibility for the times when people really are acting in bad faith. That may be in this case, but that may not be, and I don't like guilty-until-providen-innocent no matter whom it's leveled against (even a Federal agency of which I'm not the world's biggest fan, and a database the likes of which make me think we need a constitutional amendment to protect the privacy of the citizens).

mambodog 10y ago
Even if it was no-one's intent that the system is racially biased, allowing such a system to continue operate after learning of such a bias is a racist act.
- lgessler 10y ago
  I'm inclined to agree with you, but what would you say if it turned out that after much effort the engineering team was unable to make performance equal across all skin tones?
  - jacquesm 10y ago
    You mean that it would work better with darker skin?
    - handedness 10y ago
      What if the technical challenges of identifying a person's face mean it will always be challenging to identify someone when the light and/or contrast levels are lower? What then? Manually cripple the system and misidentify white suspects, too? This post's comments are a great example of why technologists are fairly terrible at engineering social solutions, much the same way sociologists tend to be terrible at engineering technical solutions.
- handedness 10y ago
  Ignoring known technical deficiencies in the system and assigning the same confidence levels in the output regardless of race could work out to be pretty racist, but what are they supposed to do, manually cripple the accuracy of the system down to the lowest common denominator? In the medical device I mentioned above, it was medically useful for caucasians and lighter-skinned Hispanics, so-so with Asians, and almost useless for Africans. So, what–don't go to market until the technology has parity across all skin tones? Not everything is about intent.
  - bilbo0s 10y ago
    I hate to be a pedant here... but I actually wrote a pretty successful 3d radiology/cardiology package. So I understand medical devices. I just wanted to hop in here and say that there is NO WAY the FDA would give you clearance to use the device you describe on African Americans under the circumstances that your posts have postulated. It seems a no brainer that the facial recognition software should not be used on African Americans either. At least, if it was MEDICAL software... it would DEFINITELY NOT be cleared with that kind of a failure rate. And any provider or delivery network using the facial recognition software on African Americans would be on his or her way to prison. But the regulations on legal software may be more lackadaisical than the regulations on medical software. Getting things exactly right may not be as important. And the penalties for misuse of medical devices are, without question, far more draconian.
    - handedness 10y ago
      > I just wanted to hop in here and say that there is NO WAY the FDA would give you clearance to use the device you describe on African Americans under the circumstances that your posts have postulated. The FDA has approved drugs that are specifically intended to target medical issues in specific ethnicities, so I'm not sure about your assessment. That said, this product's market was outside the scope of the FDA, so it's irrelevant. > It seems a no brainer that the facial recognition software should not be used on African Americans either. Any data source which produces data that's better than a random guess should be considered when trying to solve a problem. That it offends some people's social sensibilities is a separate issue. Look, I don't like the mere existence of the FBI's system, and I don't think the public's continued lack of concern for placing constitutional restraints on systems like it is a wise course forward, but to say it's a "no-brainer" when it might produce valuable leads is to focus on one aspect of the problem without considering the rest of it. Just as it would also be a mistake to consider it more accurate than it is when trying to identify dark-skinned suspects. Witness accuracy is not a new problem. It's also worth noting that the human eye is just as bad at facial recognition when the light and/or contrast is lower.

chillacy 10y ago

Depends on the camera settings too. Older movies with mostly white casts messed this up a lot, by not properly calibrating for the darker skinned characters. Conversely, movies with all black casts get this right: they tune their camera/color settings to make the faces more distinguishable. I ran into this a few weeks ago where I took a picture of a couple friends. The camera decided to pick the friend with dark skin, resulting in him looking great, and my white friend (who isn't even super pale or anything) getting washed out.

sundvor 10y ago
Yeah, I shoot semi-manual on my old D700 SLR. If you want high contrast ratios in dark areas of an image, everything else will be blown out - and you'll have to use a longer shutter speed - or suffer iso degradation or blur.
- chillacy 10y ago
  > If you want high contrast ratios in dark areas of an image, everything else will be blown out Conversely could you say that in order to capture subjects with lighter skin, you need to increase lighting in the background or else the background will clip? Here's an article about the cinematography and technology choices: https://www.washingtonpost.com/entertainment/movies/12-years-a-slave-mother-of-george-and-the-aesthetic-politics-of-filming-black-skin/2013/10/17/282af868-35cd-11e3-80c6-7e6dd8d22d8f_story.html https://www.washingtonpost.com/entertainment/movies/12-years... > he was able to photograph black actors of dramatically different skin tones in a nighttime interior scene using just everyday house lamps, thanks to a sophisticated digital camera. “I just changed the wattage of the bulb, used a dimmer, and I didn’t have to use any film lights. That kind of blew me away,” Patterson says. “The camera was able to hold both of them during the scene without any issues.” It seems like there might have been tech limitations initially, but with today's technology you can light your backgrounds and expose for either dark or light subjects well. The big challenge is having a photo with 2 subjects in the same exposure who you want to both look good. Edit: by the way I too have an old Nikon camera. I usually shoot in Aperture priority with some intentional underexposure, and then edit it up in post :)

ma2rten 10y ago

If it's a machine learning system (which virtually all face recognition systems are as far as I know) it depends more than anything on the training data. It's just the same as with people: https://en.wikipedia.org/wiki/Cross-race_effect https://en.wikipedia.org/wiki/Cross-race_effect