3 ms·

Google's AI can now identify faces from heavily pixellated images

14 points by rfreytag 9y ago

chmod775 9y ago
It absolutely can't identify faces. That's not even what it's intended to do. All it can do is create a believable high resolution face for a given pixelated one.
- GuiA 9y ago
  > All it can do is create a believable high resolution face for a given pixelated one. Yes; and, it should be emphasized, the face it will generate has no guarantee whatsoever of matching the original one, except through their heavily downscaled version.
- felixge 9y ago
  Indeed. The actual story behind this research can be found here: http://tinyclouds.org/residency/ http://tinyclouds.org/residency/ . JavaScript folks should also recognize the name of the researcher ;).
GuiA 9y ago
In the future, with further development, similar systems could be developed to add detail to pictures and video that are low resolution. One current category that lends itself to this is blurry CCTV images. This is just statistical extrapolation pushed to the absurd, and various civil rights disasters waiting to happen. Look at the outputs compared to the ground truth - the faces look absolutely nothing alike. It should be a real concern that some will see these systems as magical - being able to recover information that just isn't there - and advocate for their use in real life situations, such as security cameras as mentioned in the article. How long until an innocent person gets prison for life or the death sentence because a neural net generated a face that looks closely like theirs out of blurry footage? This is basically a modern polygraph in the making - "garbage in, garbage out" framed as science.
- rando444 9y ago
  While I agree with your statement, I was actually very impressed to see the level of detail they were able to achieve with this technology.
- seanmcdirmid 9y ago
  Reminds me of Red Dwarf’s satirical uncrop, zoom, enhance skit: http://www.dailymotion.com/video/x2qlmuy http://www.dailymotion.com/video/x2qlmuy
ibdf 9y ago
Ah-ha! TV shows will no longer have to fake it. Too bad the news is old and the title is all wrong.
sharemywin 9y ago
As with a lot the articles, I wonder what the examples they don't show look like.
ageitgey 9y ago
This is very poor reporting. The paper has nothing to do with identifying who is in a low-res image. It's about scaling up low-res images by totally making up the missing parts based on a statistical model. I guarantee the authors of this paper would be horrified by this article and how their work was mis-represented. The first paragraph of the research paper literally says "We demonstrate that this model produces a diversity of plausible high resolution images" The part at the end of the story about this being applied to CCTV images is especially misguided. But the bigger lesson here to all researchers is that they need to be aware that this is likely to happen with their work! They need to make sure they are communicating their work as clearly as possible to reduce the risk of this happening (maybe by posting a lay person-readable blog post to go with the paper that clearly says what it does and doesn't do).
- deckarep 9y ago
  Just glancing at the white paper examples shows that this article is completely sensational and missing the point. Good call out.

14 points by rfreytag 9y ago

It absolutely can't identify faces. That's not even what it's intended to do. All it can do is create a believable high resolution face for a given pixelated one.

GuiA 9y ago
> All it can do is create a believable high resolution face for a given pixelated one. Yes; and, it should be emphasized, the face it will generate has no guarantee whatsoever of matching the original one, except through their heavily downscaled version.
felixge 9y ago
Indeed. The actual story behind this research can be found here: http://tinyclouds.org/residency/ http://tinyclouds.org/residency/ . JavaScript folks should also recognize the name of the researcher ;).

GuiA 9y ago

In the future, with further development, similar systems could be developed to add detail to pictures and video that are low resolution. One current category that lends itself to this is blurry CCTV images. This is just statistical extrapolation pushed to the absurd, and various civil rights disasters waiting to happen. Look at the outputs compared to the ground truth - the faces look absolutely nothing alike. It should be a real concern that some will see these systems as magical - being able to recover information that just isn't there - and advocate for their use in real life situations, such as security cameras as mentioned in the article. How long until an innocent person gets prison for life or the death sentence because a neural net generated a face that looks closely like theirs out of blurry footage? This is basically a modern polygraph in the making - "garbage in, garbage out" framed as science.

rando444 9y ago
While I agree with your statement, I was actually very impressed to see the level of detail they were able to achieve with this technology.
seanmcdirmid 9y ago
Reminds me of Red Dwarf’s satirical uncrop, zoom, enhance skit: http://www.dailymotion.com/video/x2qlmuy http://www.dailymotion.com/video/x2qlmuy

ibdf 9y ago

Ah-ha! TV shows will no longer have to fake it. Too bad the news is old and the title is all wrong.

sharemywin 9y ago

As with a lot the articles, I wonder what the examples they don't show look like.

ageitgey 9y ago

This is very poor reporting. The paper has nothing to do with identifying who is in a low-res image. It's about scaling up low-res images by totally making up the missing parts based on a statistical model. I guarantee the authors of this paper would be horrified by this article and how their work was mis-represented. The first paragraph of the research paper literally says "We demonstrate that this model produces a diversity of plausible high resolution images" The part at the end of the story about this being applied to CCTV images is especially misguided. But the bigger lesson here to all researchers is that they need to be aware that this is likely to happen with their work! They need to make sure they are communicating their work as clearly as possible to reduce the risk of this happening (maybe by posting a lay person-readable blog post to go with the paper that clearly says what it does and doesn't do).

deckarep 9y ago
Just glancing at the white paper examples shows that this article is completely sensational and missing the point. Good call out.