5 ms·

The whole article seems to be a marketing post. I completely agree that you could build a face recognition system in 60$ or less, however the image match descr

The whole article seems to be a marketing post. I completely agree that you could build a face recognition system in 60$ or less, however the image match described in the post is unlikely to be the real system results, I've been studying image recognition for a year now and it is almost impossible to train a model without having thousands of images classified somehow (usually by hands). One of the biggest problems in AI is that NN cannot match an object in 2 images if it was taken from different perspective unless R. Madonna was hanging out there for 9 hours to be filmed from different angles to collect enough of photos from camera (not taking into account different image resolutions and camera quality) Thanks for the article, however, the myth is BUSTED :D

mhb 7y ago
That sounds reasonable but what do you think they're marketing?
- tarasmatsyk 7y ago
  depends on nytimes KPIs :D maybe just growing click rates on the article
shraremywin2 7y ago
You just training on do two face images come from the same person. https://towardsdatascience.com/one-shot-learning-face-recognition-using-siamese-neural-network-a13dcf739e https://towardsdatascience.com/one-shot-learning-face-recogn...
Yuval_Halevi 7y ago
99% of the stuff you see online is marketing. Someone want to benefit from your visit. If they don't try to sell you anything, they want to get traffic so they will create an attractive clickbait title
- matt4077 7y ago
  Well, that’s a conveniently tautological conspiracy theory.
dalbasal 7y ago
Sure, but that's not that big a barrier. First, there are some pretty good classified mugshot datasets in existence. FB's, for one. Second, you can get some mileage out of just matching images to eachother. One or more cameras track an individual to get multiple angles. You may not know the individual's name, but you will know that it's the same person walking past the cameras every day. Third, you can always target a limited set of faces. Don't casinos use something like this for persona non grata?
nickelcitymario 7y ago
The article may very well have exaggerated how easy it is to pull off. However, I would actually take it a step further. I could build a facial recognition system for $0 using open-source software and freely available camera feeds. That's the real crux of the article: The data (camera feeds) is out there. The software is out there. Anyone could do this. Therefore, what are we doing to prevent it, if anything?
taneq 7y ago
You're talking about training data, but (as I understand it) current commercial facial recognition may use learned features but from there it uses traditional biometrics to identify faces. So this criticism doesn't necessarily hold as these biometrics could easily be one-shot. (Any facial recognition gurus feel free to enlighten me, of course! :)
ratsimihah 7y ago
Right on, for a minute I was wondering if there hasn't been a leap in the efficiency of neural nets that I missed.
thinkcontext 7y ago
Did you read the article? They show the image that was matched of the professor compared to the training image, they are completely different perspectives.
mabbo 7y ago
> One of the biggest problems in AI is that NN cannot match an object in 2 images if it was taken from different perspective I think you've missed the last 5-ish years of NN development. Some of the latest Face2Vec stuff can handle different perspectives pretty well by either building a system that can predict 3D representations of the face from one shot[0] or by just training on two different perspectives of the same face and having the system output the same vector (I've read some work on this but Google is failing me at finding a reference). If you can recognize people from different angles, computers will eventually be able to do it better than you. Probably already. [0]https://arxiv.org/pdf/1703.10714.pdf https://arxiv.org/pdf/1703.10714.pdf
- bsenftner 7y ago
  Bingo. 1) Train an algorithm that given any angle view of a human head, produce a 3D reconstruction of said head. (Depending upon the amount of face visible, the accuracy of the reconstruction increases. 2) Train an algorithm that given a human head geometry, it identifies any deformation due to facial expression and neutralizes that expression. 3) Trivial algorithm: given a human head geometry not facing the camera, rotate that head such that it is facing the camera. 4) render the human head geometry without expression, facing the camera, using the video image source as the texture map for said head. 5) Perform FR.
villgax 7y ago
You've been out of the loop, just look at FaceNet, your face in any orientation for a human to recognise is enough for the software to do so as well. Is already open source & doesn't need any manual tuning after it has been trained, also scales excellently.

mhb 7y ago

That sounds reasonable but what do you think they're marketing?

tarasmatsyk 7y ago
depends on nytimes KPIs :D maybe just growing click rates on the article

shraremywin2 7y ago

You just training on do two face images come from the same person. https://towardsdatascience.com/one-shot-learning-face-recognition-using-siamese-neural-network-a13dcf739e https://towardsdatascience.com/one-shot-learning-face-recogn...

Yuval_Halevi 7y ago

99% of the stuff you see online is marketing. Someone want to benefit from your visit. If they don't try to sell you anything, they want to get traffic so they will create an attractive clickbait title

matt4077 7y ago
Well, that’s a conveniently tautological conspiracy theory.

dalbasal 7y ago

Sure, but that's not that big a barrier. First, there are some pretty good classified mugshot datasets in existence. FB's, for one. Second, you can get some mileage out of just matching images to eachother. One or more cameras track an individual to get multiple angles. You may not know the individual's name, but you will know that it's the same person walking past the cameras every day. Third, you can always target a limited set of faces. Don't casinos use something like this for persona non grata?

nickelcitymario 7y ago

The article may very well have exaggerated how easy it is to pull off. However, I would actually take it a step further. I could build a facial recognition system for $0 using open-source software and freely available camera feeds. That's the real crux of the article: The data (camera feeds) is out there. The software is out there. Anyone could do this. Therefore, what are we doing to prevent it, if anything?

taneq 7y ago

You're talking about training data, but (as I understand it) current commercial facial recognition may use learned features but from there it uses traditional biometrics to identify faces. So this criticism doesn't necessarily hold as these biometrics could easily be one-shot. (Any facial recognition gurus feel free to enlighten me, of course! :)

ratsimihah 7y ago

Right on, for a minute I was wondering if there hasn't been a leap in the efficiency of neural nets that I missed.

thinkcontext 7y ago

Did you read the article? They show the image that was matched of the professor compared to the training image, they are completely different perspectives.

mabbo 7y ago

> One of the biggest problems in AI is that NN cannot match an object in 2 images if it was taken from different perspective I think you've missed the last 5-ish years of NN development. Some of the latest Face2Vec stuff can handle different perspectives pretty well by either building a system that can predict 3D representations of the face from one shot[0] or by just training on two different perspectives of the same face and having the system output the same vector (I've read some work on this but Google is failing me at finding a reference). If you can recognize people from different angles, computers will eventually be able to do it better than you. Probably already. [0]https://arxiv.org/pdf/1703.10714.pdf https://arxiv.org/pdf/1703.10714.pdf

bsenftner 7y ago
Bingo. 1) Train an algorithm that given any angle view of a human head, produce a 3D reconstruction of said head. (Depending upon the amount of face visible, the accuracy of the reconstruction increases. 2) Train an algorithm that given a human head geometry, it identifies any deformation due to facial expression and neutralizes that expression. 3) Trivial algorithm: given a human head geometry not facing the camera, rotate that head such that it is facing the camera. 4) render the human head geometry without expression, facing the camera, using the video image source as the texture map for said head. 5) Perform FR.

villgax 7y ago

You've been out of the loop, just look at FaceNet, your face in any orientation for a human to recognise is enough for the software to do so as well. Is already open source & doesn't need any manual tuning after it has been trained, also scales excellently.