3 ms·

Smaller neural nets are commonly used in character recognition, but typical smart cameras or embedded robot controllers don't have anywhere near enough compute

Smaller neural nets are commonly used in character recognition, but typical smart cameras or embedded robot controllers don't have anywhere near enough compute to run deep networks in real time. The Fanuc I was ranting about in this thread had something like 64MB of RAM in 2018. Maybe some systems are out there using Coral TPUs with custom TF Lite models. As for PC-based systems, I would be very surprised if deep learning models weren't being used in production somewhere. But in a factory environment you can go a very long way with primitive feature recognition and good control over the scene and lighting, and the customer just cares that whatever you're doing just works and any new method will have to be enough of an improvement to be worth the cost of development time.

vidanay 4y ago
The 900lb gorilla in the deep learning room that everyone likes to ignore is that machine learning is horrible at providing corrective action data. Traditional machine vision is well adapted to providing statistical data such as "the diameter of the pizza is out of tolerance by 8mm" or "there are supposed to be 22 pepperonis on the pizza, but only 19 were found". Machine learning leans towards "it's not a good pizza" and doesn't provide a lot of additional data.
- krisoft 4y ago
  > machine learning is horrible at providing corrective action data I don’t recognise the truth in what you are writing. > there are supposed to be 22 pepperonis on the pizza, but only 19 were found Instance segmentation is a solved problem. A properly constructed and trained neural network can tell you exactly how many pepperonies it sees and exactly where. Telling if that is the right number is a trivial problem from there. > the diameter of the pizza is out of tolerance by 8mm Here too, the neural network can recognise the edges of the pizza and then you can fit a shape to it. You can do this second step either with classical algorithms or with a machine learning one. (I would use a classical algorithm if the pizza is meant to be circular or rectangular shaped, and a machine learning algorithm if they are aiming for something weird, like an Italy shaped pizza or something.) > Machine learning leans towards "it's not a good pizza" Sounds like you have only heard of simple classifier models.
  - vidanay 4y ago
    I will accept your opinion as I have never implemented a complete ML based solution. All of my opinion is based on promises and demonstrations for ML products such as Cognex VIDI. If those systems have capabilities like you describe, they have not been well presented during their sales pitches.
    - version_five 4y ago
      I've observed that the ML (or otherwise) model quality is not really the bottleneck in most computer vision systems so it gets less attention. The state of the art is way beyond what you would see in an implemented solution (like cognex), but some combination of market immaturity and that not being the biggest problem means there's not much industry demand for really good models
    - krisoft 4y ago
      > All of my opinion is based on promises and demonstrations for ML products such as Cognex VIDI. I see! Thank you for the explanation. That now makes sense. Basically you were talking about what is available on the market as a product, and I was talking about what the state of the art is in machine learning. Now obviously if you actually want to put a factory line together you care about the available products, not what is possible in theory. It is kind of like asking someone if it is possible to travel to the moon. If you are asking a physicist they will do some calculations with the rocket equation and will tell you that it is perfectly possible. If you ask the same question from a travel agent they will tell you it is not possible because they can’t sell you a moon holiday right there and then. They are both right, just in different contexts. > If those systems have capabilities like you describe, they have not been well presented during their sales pitches. All I can tell to those companies is that they should “git gud”. :) Thank you for your explanation about the context you were talking about!
- jeffreyrogers 4y ago
  There's a lot of progress on this recently with things like conformal prediction.
- version_five 4y ago
  As the other person said, this is largely a strawman. It would be possible to build some kind of black box system, but it's not mandatory. An OK/NOK pizza classifier could do so without telling you why, and in some measures may be more robust than some kind of filter, threshold, morphological rules, pepperoni counter, but the latter will not tell you if something is wrong with the crust. A pepperoni object detector would be trivial and way more robust than whatever classical pepperoni finder you could build.
fest 4y ago
> As for PC-based systems, I would be very surprised if deep learning models weren't being used in production somewhere. They definitely are. ~5 years ago I built a PC-based system that detected grain direction of wooden boards (looking at the end of the board). Initially I resisted the ML approaches and my first attempt was basically hand-crafted image analysis pipeline- split the image in segments, apply Gabor filter with kernels of various angles and try to fit a curve to results. It kind-of-worked but I wasn't entirely happy with it's performance on the test data. Even the simple classifier models that could execute on a fanless PC without a GPU outperformed my solution, and after a few more training runs the handcrafted code was replaced by #include <tensorflow.h>. This year I'll have to extend the system with on-site training mode, where an operator has a pushbutton to label the images and re-train the model.
fest 4y ago
Also, I'm pretty sure all the major smart camera vendors have projects underway which utilize NVidia Jetson.

vidanay 4y ago

The 900lb gorilla in the deep learning room that everyone likes to ignore is that machine learning is horrible at providing corrective action data. Traditional machine vision is well adapted to providing statistical data such as "the diameter of the pizza is out of tolerance by 8mm" or "there are supposed to be 22 pepperonis on the pizza, but only 19 were found". Machine learning leans towards "it's not a good pizza" and doesn't provide a lot of additional data.

krisoft 4y ago
> machine learning is horrible at providing corrective action data I don’t recognise the truth in what you are writing. > there are supposed to be 22 pepperonis on the pizza, but only 19 were found Instance segmentation is a solved problem. A properly constructed and trained neural network can tell you exactly how many pepperonies it sees and exactly where. Telling if that is the right number is a trivial problem from there. > the diameter of the pizza is out of tolerance by 8mm Here too, the neural network can recognise the edges of the pizza and then you can fit a shape to it. You can do this second step either with classical algorithms or with a machine learning one. (I would use a classical algorithm if the pizza is meant to be circular or rectangular shaped, and a machine learning algorithm if they are aiming for something weird, like an Italy shaped pizza or something.) > Machine learning leans towards "it's not a good pizza" Sounds like you have only heard of simple classifier models.
- vidanay 4y ago
  I will accept your opinion as I have never implemented a complete ML based solution. All of my opinion is based on promises and demonstrations for ML products such as Cognex VIDI. If those systems have capabilities like you describe, they have not been well presented during their sales pitches.
  - version_five 4y ago
    I've observed that the ML (or otherwise) model quality is not really the bottleneck in most computer vision systems so it gets less attention. The state of the art is way beyond what you would see in an implemented solution (like cognex), but some combination of market immaturity and that not being the biggest problem means there's not much industry demand for really good models
  - krisoft 4y ago
    > All of my opinion is based on promises and demonstrations for ML products such as Cognex VIDI. I see! Thank you for the explanation. That now makes sense. Basically you were talking about what is available on the market as a product, and I was talking about what the state of the art is in machine learning. Now obviously if you actually want to put a factory line together you care about the available products, not what is possible in theory. It is kind of like asking someone if it is possible to travel to the moon. If you are asking a physicist they will do some calculations with the rocket equation and will tell you that it is perfectly possible. If you ask the same question from a travel agent they will tell you it is not possible because they can’t sell you a moon holiday right there and then. They are both right, just in different contexts. > If those systems have capabilities like you describe, they have not been well presented during their sales pitches. All I can tell to those companies is that they should “git gud”. :) Thank you for your explanation about the context you were talking about!
jeffreyrogers 4y ago
There's a lot of progress on this recently with things like conformal prediction.
version_five 4y ago
As the other person said, this is largely a strawman. It would be possible to build some kind of black box system, but it's not mandatory. An OK/NOK pizza classifier could do so without telling you why, and in some measures may be more robust than some kind of filter, threshold, morphological rules, pepperoni counter, but the latter will not tell you if something is wrong with the crust. A pepperoni object detector would be trivial and way more robust than whatever classical pepperoni finder you could build.

fest 4y ago

> As for PC-based systems, I would be very surprised if deep learning models weren't being used in production somewhere. They definitely are. ~5 years ago I built a PC-based system that detected grain direction of wooden boards (looking at the end of the board). Initially I resisted the ML approaches and my first attempt was basically hand-crafted image analysis pipeline- split the image in segments, apply Gabor filter with kernels of various angles and try to fit a curve to results. It kind-of-worked but I wasn't entirely happy with it's performance on the test data. Even the simple classifier models that could execute on a fanless PC without a GPU outperformed my solution, and after a few more training runs the handcrafted code was replaced by #include <tensorflow.h>. This year I'll have to extend the system with on-site training mode, where an operator has a pushbutton to label the images and re-train the model.

fest 4y ago

Also, I'm pretty sure all the major smart camera vendors have projects underway which utilize NVidia Jetson.