8 ms·

What if you take a picture of a modified picture?

jack_pp 4y ago
I like this, if we can fool humans with VR goggles then surely we can fool the camera by using a sufficiently detailed print or a high resolution display
- CharlesW 4y ago
  A more effective version of this would capture a 3D depth map with the 2D image.
ChrisLomont 4y ago
You'll never remove aliasing artifacts.
- TeeMassive 4y ago
  Not all fake images are images taken from cameras (e.g. CGI, AI)
- josephcsible 4y ago
  ILM figured out how to: https://en.wikipedia.org/wiki/StageCraft https://en.wikipedia.org/wiki/StageCraft
  - ChrisLomont 4y ago
    What in that makes you think there are no aliasing artifacts? Yes, you can take a movie of a movie, and you introduce artifacts that simple frequency analysis will find. There's no reason ILM would even try to remove this for such an application - they simply want pretty visuals, not mathematically indistinguishable from reality digital compositing. Take a photo. It's not a grid of values. You already screwed with high frequency components. Take a photo of that - you changed these again. It's not hard to detect this. There's an entire field for detecting forensically altered things, and no current tech goes undetected, certainly not stuff simply made for movies. Or provide some link, paper, or serious scientific claim that ILM has magically removed all aliasing from images, which would, well, simply violate physics. What they actually do is try to move aliasing to less visible portions of a frequency spectrum - but those signals are still there and detectable.
    - praash 4y ago
      > -- would, well, simply violate physics. Nope. Aliasing can be eliminated by low-pass filtering (blurring) the input signal before sampling it. This could be done by keeping the background even slightly out of focus, or coating the display. Unless it's traded away for performance (or style choice), you are experiencing alias-free sound and image processing all day long.
      - ChrisLomont 4y ago
        Yep. >Aliasing can be eliminated by low-pass filtering (blurring) the input signal before sampling it. Nope. I have a PhD in math, and have been doing filtering stuff for decades. You do not eliminate aliasing, you push the error out of the visible spectrum - end of story. The naive, physically impossible undergrad textbooks might say differently, just like they'd teach Newton as gravity and ignore relativity, or teach the ideal gas law and ignore that real gasses are vastly more complex in behavior. This is Wikipedia level knowledge: from the low-pass filter page [1]: "However, the ideal filter is impossible to realize without also having signals of infinite extent in time, and so generally needs to be approximated for real ongoing signals, because the sinc function's support region extends to all past and future times." So no ideal filter, sorry. That means leakage, which happens in every part of engineering (and is forced by physics, despite you not wanting to believe it). Ideal filters are math, not physics. And this level is vastly below what is done in industry, what theory (and physics) forces onto real world filtering. Once you discretize the signal, even after filtering, you have introduce lots of noise in lots of frequency ranges. The simple reason is that you are forced to perturb real sample values (which would require infinitely precise measurements) to the discrete ones your sensor records. These perturbations do not fit your frequency requirements. Don't believe it? Take a photo, apply your "ideal" low pass filter, then fit the resulting image using sin and cos waves of frequency no higher than than your filter - and you'll see it's impossible. You can only approximate it - and that is empirical proof you can do at home. Another entire level that undergrad level understandings miss is that pretty much zero of the theorems you think apply actually do - because pretty much every one uses infinite precision samples (Nyquist, for example), infinite support reconstruction (physically impossible), and so on. In practice all of these things leave noise in various parts of frequency spectrums. So, still want to argue that filters would somehow remove aliasing issues? Especially over a sensor with the number of samples and precision that a modern Sony sensor has? >you are experiencing alias-free sound and image processing all day long You are confusing Nyquist with reality - and ignoring that Nyquist doesn't even apply to audio (or video) - it's only an approximation to what is really happening. As an approximation it's useful, but the requirements of the theorem are not met in reality and the reconstruction is not met in reality. Don't believe me? Go look at the proof for the Nyquist theorem, and pay attention to all the requirements that no physical device can meet. it's the frictionless spherical cow of audio processing. One glaring area it fails is the need for the infinite support sinc filter - which again is why real engineering is not so simplistic. [1] https://en.wikipedia.org/wiki/Low-pass_filter https://en.wikipedia.org/wiki/Low-pass_filter
TeeMassive 4y ago
Then you have a signed picture of a modified picture which can't be altered further. Cryptographic signatures only ensures data integrity and authenticity after the cryptographic signature was made.
- remram 4y ago
  It can totally be altered further. Just print the altered version and take a picture again.
  - TeeMassive 4y ago
    Not without having a valid signature
    - remram 4y ago
      Do you mean "not with"? The camera generates a valid signature every time you take a picture...
Rackedup 4y ago
reminds me of NFTs... go Sony...

I like this, if we can fool humans with VR goggles then surely we can fool the camera by using a sufficiently detailed print or a high resolution display

CharlesW 4y ago
A more effective version of this would capture a 3D depth map with the 2D image.

ChrisLomont 4y ago

You'll never remove aliasing artifacts.

TeeMassive 4y ago
Not all fake images are images taken from cameras (e.g. CGI, AI)
josephcsible 4y ago
ILM figured out how to: https://en.wikipedia.org/wiki/StageCraft https://en.wikipedia.org/wiki/StageCraft
- ChrisLomont 4y ago
  What in that makes you think there are no aliasing artifacts? Yes, you can take a movie of a movie, and you introduce artifacts that simple frequency analysis will find. There's no reason ILM would even try to remove this for such an application - they simply want pretty visuals, not mathematically indistinguishable from reality digital compositing. Take a photo. It's not a grid of values. You already screwed with high frequency components. Take a photo of that - you changed these again. It's not hard to detect this. There's an entire field for detecting forensically altered things, and no current tech goes undetected, certainly not stuff simply made for movies. Or provide some link, paper, or serious scientific claim that ILM has magically removed all aliasing from images, which would, well, simply violate physics. What they actually do is try to move aliasing to less visible portions of a frequency spectrum - but those signals are still there and detectable.
  - praash 4y ago
    > -- would, well, simply violate physics. Nope. Aliasing can be eliminated by low-pass filtering (blurring) the input signal before sampling it. This could be done by keeping the background even slightly out of focus, or coating the display. Unless it's traded away for performance (or style choice), you are experiencing alias-free sound and image processing all day long.
    - ChrisLomont 4y ago
      Yep. >Aliasing can be eliminated by low-pass filtering (blurring) the input signal before sampling it. Nope. I have a PhD in math, and have been doing filtering stuff for decades. You do not eliminate aliasing, you push the error out of the visible spectrum - end of story. The naive, physically impossible undergrad textbooks might say differently, just like they'd teach Newton as gravity and ignore relativity, or teach the ideal gas law and ignore that real gasses are vastly more complex in behavior. This is Wikipedia level knowledge: from the low-pass filter page [1]: "However, the ideal filter is impossible to realize without also having signals of infinite extent in time, and so generally needs to be approximated for real ongoing signals, because the sinc function's support region extends to all past and future times." So no ideal filter, sorry. That means leakage, which happens in every part of engineering (and is forced by physics, despite you not wanting to believe it). Ideal filters are math, not physics. And this level is vastly below what is done in industry, what theory (and physics) forces onto real world filtering. Once you discretize the signal, even after filtering, you have introduce lots of noise in lots of frequency ranges. The simple reason is that you are forced to perturb real sample values (which would require infinitely precise measurements) to the discrete ones your sensor records. These perturbations do not fit your frequency requirements. Don't believe it? Take a photo, apply your "ideal" low pass filter, then fit the resulting image using sin and cos waves of frequency no higher than than your filter - and you'll see it's impossible. You can only approximate it - and that is empirical proof you can do at home. Another entire level that undergrad level understandings miss is that pretty much zero of the theorems you think apply actually do - because pretty much every one uses infinite precision samples (Nyquist, for example), infinite support reconstruction (physically impossible), and so on. In practice all of these things leave noise in various parts of frequency spectrums. So, still want to argue that filters would somehow remove aliasing issues? Especially over a sensor with the number of samples and precision that a modern Sony sensor has? >you are experiencing alias-free sound and image processing all day long You are confusing Nyquist with reality - and ignoring that Nyquist doesn't even apply to audio (or video) - it's only an approximation to what is really happening. As an approximation it's useful, but the requirements of the theorem are not met in reality and the reconstruction is not met in reality. Don't believe me? Go look at the proof for the Nyquist theorem, and pay attention to all the requirements that no physical device can meet. it's the frictionless spherical cow of audio processing. One glaring area it fails is the need for the infinite support sinc filter - which again is why real engineering is not so simplistic. [1] https://en.wikipedia.org/wiki/Low-pass_filter https://en.wikipedia.org/wiki/Low-pass_filter

TeeMassive 4y ago

Then you have a signed picture of a modified picture which can't be altered further. Cryptographic signatures only ensures data integrity and authenticity after the cryptographic signature was made.

remram 4y ago
It can totally be altered further. Just print the altered version and take a picture again.
- TeeMassive 4y ago
  Not without having a valid signature
  - remram 4y ago
    Do you mean "not with"? The camera generates a valid signature every time you take a picture...

Rackedup 4y ago

reminds me of NFTs... go Sony...