K-3 III - what is "deep learning with learning promises" in AF-system?

Grokh · 08-21-2021, 02:31 PM

Dear colleagues, what does the phrase in the description of the camera on Ricoh's website mean, which in translation sounds like this: "An overview of predictive autofocus algorithms and the ability to recognize faces, eyes and objects such as birds, through deep learning with training, promises to improve AF-C performance in a number of situations "?

vonBaloney · 08-21-2021, 02:41 PM

deep learning = neural networks (a currently dominant machine learning algorithm for anything image related)

Modern neural networks have gotten really good at recognizing objects -- not just faces anymore. So in the future, so you might have a special "bird in flight" AF mode that would actually recognize the bird and be able to focus on the eyes, for instance.

vector · 08-21-2021, 02:44 PM - **2 Likes**

At a high level they create computer program to analyze images and then feed it tons of images of a certain type, say a bird in flight. This is used to train the software to recognize the shape of a bird in flight. Once the software is trained it can be used in a camera to recognize a shape rather than just detecting contrast. This is often referred to as artificial intelligence or deep learning, but it's really just someone writing some capable software and tweaking it to recognize patterns.

**PintofDog** · 08-21-2021, 02:50 PM - **4 Likes**

Frankly it’s marketing. I work in the field currently know as artificial intelligence and terms are bandied about to suit.
Deep learning generally refers to the use of neural networks with multiple layers. These have always been known about but up until now the hardware required was not available, today it is. The training is an interesting part, we only use supervised training, where we decide what the system is trained on. Unsupervised training allows the AI system to make its own decisions. I have heard many systems claim their technology does this but they soon fade away. Regardless it is the way forward.

Grokh · 08-21-2021, 02:52 PM

Thank you, colleagues.
Let me clarify the question a little: is it possible to conclude that this system is self-learning?
Or is it just that she is "pretrained"?

vonBaloney · 08-21-2021, 02:54 PM - **2 Likes**

Originally posted by Grokh

Thank you, colleagues.
Let me clarify the question a little: is it possible to conclude that this system is self-learning?
Or is it just that she is "pretrained"?

100% pretrained unless they say otherwise, which would be VERY surprising (computing power needed for training is immense). It could get updated models with firmware updates though (which is quite likely). I could even see them having a separate "model update" apart from the main firmware for that. But for a first version of this sort of thing there is no way they'd be jumping straight to self-learning. (Come back in 10 years for that maybe.)

Grokh · 08-21-2021, 02:59 PM

Originally posted by vonBaloney

100% pretrained unless they say otherwise, which would be VERY surprising (computing power needed for training is immense). It could get updated models with firmware updates though (which is quite likely). I could even see them having a separate "model update" apart from the main firmware for that. But for a first version of this sort of thing there is no way they'd be jumping straight to self-learning. (Come back in 10 years for that maybe.)

Thank you, I thought the processing power was too low, but suddenly ...

AfterPentax Mark II · 08-21-2021, 03:02 PM

It is selflearning and that with the lapse of time that you take pictures (the more the better) with the K-3 Mark III, it recognizes the type of objects that you take pictures of. It is in this way that the predicative autofocus is trained to recognize your objects (read humans, birds, animals and so on) and it is used by the "predictive autofocus algorithms" as you put it. It means that the more you take pictures of birds or animals and other moving objects the better the respons of the autofocus with AF.C autofocus, because it learned to recognize the moving objects that you have taken pictures of. Hope this helps.
It is no marketing, but we all know that Japanese manufacturers (and many others) like to use difficult words to impress us, but it is really a simple learning process that is used by Honeywell, to mention a company, for adapting your central heating to your way of life and to weather conditions to start your central heating so that at a certain time, usually the time you wake up, the house is warm in the way you like it.
This system of recognizing objects is also found in some photo processing programs that sort your pictures of your cats, your dogs your family in these specific categories if you want to. It also means that with the laps of time your camera will be responding faster. Deep learning, just sounds goods but what they meant is that it will take (a long) time

.

Kunzite · 08-21-2021, 03:11 PM

Originally posted by AfterPentax Mark II

It is selflearning and that with the lapse of time that you take pictures (the more the better) with the K-3 Mark III, it recognizes the type of objects that you take pictures of. It is in this way that the predicative autofocus is trained to recognize your objects (read humans, birds, animals and so on) and it is used by the "predictive autofocus algorithms" as you put it. It means that the more you take pictures of birds or animals and other moving objects the better the respons of the autofocus with AF.C autofocus, because it learned to recognize the moving objects that you have taken pictures of. Hope this helps.

I'm quite sure it doesn't learn as you use it.
The "learning part" is very resource intensive, no camera should be capable of doing it. Possibly not even your PC.
Then, I have doubts about the idea of training it to recognize (what?) just by using the camera, and no other clue.

L.E. I see vonBaloney said something similar not long ago. Well, it's worth repeating.

AfterPentax Mark II · 08-21-2021, 03:28 PM - **1 Like**

Originally posted by Kunzite

I'm quite sure it doesn't learn as you use it.
The "learning part" is very resource intensive, no camera should be capable of doing it. Possibly not even your PC.
Then, I have doubts about the idea of training it to recognize (what?) just by using the camera, and no other clue.

L.E. I see vonBaloney said something similar not long ago. Well, it's worth repeating.

Yes it does, because it is a very simple system and the use of those words sets everybody on the wrong foot. It is maybe not "learning" but you can think of it as tagging things that share alike things. Recognizing eyes, can be a problem, because how does it know it is human or animal? They combine it with a laughing face which is mostly human. Eye and face recognition is wide spread in camera world. Microsoft Money to mention something a bit PC-like can recognize that you make the same kind of payments and then asks you if that is correct and if you want to tag it and each time such a payment is entered it will ask you if it must put an ok on it for this month. So if your camera recognizes something and focuses on that you can proceed or you can change the focus to the subject intended. You camera's programmed hardware can remember that. I edited my original post and elaborated on the subject with a few examples where it is used.

vonBaloney · 08-21-2021, 04:30 PM - **1 Like**

If it uses the phrase "deep learning" that means neural networks -- that is a very specific thing and it is not a "general" phrase or marketing-speak. And I can assure you the camera is not doing neural network training. Again, maybe someday, but no way.

cfraz · 08-21-2021, 04:50 PM

Originally posted by AfterPentax Mark II

Yes it does, because it is a very simple system and the use of those words sets everybody on the wrong foot. It is maybe not "learning" but you can think of it as tagging things that share alike things. Recognizing eyes, can be a problem, because how does it know it is human or animal? They combine it with a laughing face which is mostly human. Eye and face recognition is wide spread in camera world. Microsoft Money to mention something a bit PC-like can recognize that you make the same kind of payments and then asks you if that is correct and if you want to tag it and each time such a payment is entered it will ask you if it must put an ok on it for this month. So if your camera recognizes something and focuses on that you can proceed or you can change the focus to the subject intended. You camera's programmed hardware can remember that. I edited my original post and elaborated on the subject with a few examples where it is used.

Absolutely not.

disconnekt · 08-21-2021, 05:44 PM - **1 Like**

Originally posted by vonBaloney

100% pretrained unless they say otherwise, which would be VERY surprising (computing power needed for training is immense). It could get updated models with firmware updates though (which is quite likely). I could even see them having a separate "model update" apart from the main firmware for that. But for a first version of this sort of thing there is no way they'd be jumping straight to self-learning. (Come back in 10 years for that maybe.)

Im surprised Ricoh hasn come out & said "Hey Pentax users, share your photography photos with us so we can use the data we get from the photos to help improve the firmware we send out to make our DSLR's more accurate in various scenarios"

Cuz as I see it, getting data directly from your user base on what they shoot (where exactly to improve on) will help out alot on witu future firmware updates/DSLR releases

**jpzk** · 08-21-2021, 07:03 PM

Originally posted by AfterPentax Mark II

It is selflearning and that with the lapse of time that you take pictures (the more the better) with the K-3 Mark III, it recognizes the type of objects that you take pictures of. It is in this way that the predicative autofocus is trained to recognize your objects (read humans, birds, animals and so on) and it is used by the "predictive autofocus algorithms" as you put it. It means that the more you take pictures of birds or animals and other moving objects the better the respons of the autofocus with AF.C autofocus, because it learned to recognize the moving objects that you have taken pictures of. Hope this helps.
It is no marketing, but we all know that Japanese manufacturers (and many others) like to use difficult words to impress us, but it is really a simple learning process that is used by Honeywell, to mention a company, for adapting your central heating to your way of life and to weather conditions to start your central heating so that at a certain time, usually the time you wake up, the house is warm in the way you like it.
This system of recognizing objects is also found in some photo processing programs that sort your pictures of your cats, your dogs your family in these specific categories if you want to. It also means that with the laps of time your camera will be responding faster. Deep learning, just sounds goods but what they meant is that it will take (a long) time

.

Would the K3III be able to "learn" while in AF.S mode rather than in AF.C and "extrapolate" this learning to AF.C ?

kypfer · 08-21-2021, 10:17 PM - **5 Likes**

Sorry, this concept that a camera could be "self-learning" has got to be mis-leading at the very least!
If it were the case, within virtually no time at all every K-3iii would be "different" and in the case of needing repair, all this "learning" could be lost and the user would have to start over!
Having to re-enter carefully crafted User Modes is pain enough (why oh why can't we save and load User Modes to/from SD card?), to have to "re-train" a camera that has made several (tens of) thousand exposures could be a major problem

08-21-2021, 02:31 PM	#1
Grokh Senior Member Join Date: Feb 2011 Location: St.-Petersburg Photos: Albums Posts: 267	K-3 III - what is "deep learning with learning promises" in AF-system? Dear colleagues, what does the phrase in the description of the camera on Ricoh's website mean, which in translation sounds like this: "An overview of predictive autofocus algorithms and the ability to recognize faces, eyes and objects such as birds, through deep learning with training, promises to improve AF-C performance in a number of situations "?

08-21-2021, 02:41 PM	#2
vonBaloney Pentaxian Join Date: Oct 2011 Location: Albuquerque, NM Posts: 6,029	deep learning = neural networks (a currently dominant machine learning algorithm for anything image related) Modern neural networks have gotten really good at recognizing objects -- not just faces anymore. So in the future, so you might have a special "bird in flight" AF mode that would actually recognize the bird and be able to focus on the eyes, for instance. Last edited by vonBaloney; 08-22-2021 at 06:44 AM.

08-21-2021, 03:02 PM	#8
AfterPentax Mark II Pentaxian Join Date: Mar 2021 Posts: 1,465	It is selflearning and that with the lapse of time that you take pictures (the more the better) with the K-3 Mark III, it recognizes the type of objects that you take pictures of. It is in this way that the predicative autofocus is trained to recognize your objects (read humans, birds, animals and so on) and it is used by the "predictive autofocus algorithms" as you put it. It means that the more you take pictures of birds or animals and other moving objects the better the respons of the autofocus with AF.C autofocus, because it learned to recognize the moving objects that you have taken pictures of. Hope this helps. It is no marketing, but we all know that Japanese manufacturers (and many others) like to use difficult words to impress us, but it is really a simple learning process that is used by Honeywell, to mention a company, for adapting your central heating to your way of life and to weather conditions to start your central heating so that at a certain time, usually the time you wake up, the house is warm in the way you like it. This system of recognizing objects is also found in some photo processing programs that sort your pictures of your cats, your dogs your family in these specific categories if you want to. It also means that with the laps of time your camera will be responding faster. Deep learning, just sounds goods but what they meant is that it will take (a long) time . Last edited by AfterPentax Mark II; 08-21-2021 at 03:12 PM. Reason: additonal information added because of the other quick responses.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Abstract Deep Dream Generator - Deep Style V2	boriscleto	Post Your Photos!	13	10-25-2021 06:39 PM
Sony a6000 promises world's fastest AF and 11 fps subject tracking	jogiba	Canon, Nikon, Sony, and Other Camera Brands	51	03-28-2014 07:43 AM
Promises, promises K3 24mp New tracking AF	solar1	Pentax News and Rumors	406	10-05-2013 05:52 PM
Ricoh Imaging promises new hi-end rumours, possible another company name change	Uluru	Photographic Industry and Professionals	21	08-27-2013 04:40 AM
Nature Deep Deep Pink	eaglem	Post Your Photos!	11	05-19-2012 07:50 AM

08-21-2021, 02:44 PM - 2 Likes	#3
vector Pentaxian Join Date: May 2015 Location: Alberta Posts: 713	At a high level they create computer program to analyze images and then feed it tons of images of a certain type, say a bird in flight. This is used to train the software to recognize the shape of a bird in flight. Once the software is trained it can be used in a camera to recognize a shape rather than just detecting contrast. This is often referred to as artificial intelligence or deep learning, but it's really just someone writing some capable software and tweaking it to recognize patterns.

08-21-2021, 02:50 PM - 4 Likes	#4
PintofDog Site Supporter Join Date: Jan 2017 Posts: 50	Frankly it’s marketing. I work in the field currently know as artificial intelligence and terms are bandied about to suit. Deep learning generally refers to the use of neural networks with multiple layers. These have always been known about but up until now the hardware required was not available, today it is. The training is an interesting part, we only use supervised training, where we decide what the system is trained on. Unsupervised training allows the AI system to make its own decisions. I have heard many systems claim their technology does this but they soon fade away. Regardless it is the way forward.

08-21-2021, 02:52 PM	#5
Grokh Senior Member Join Date: Feb 2011 Location: St.-Petersburg Photos: Albums Posts: 267 Original Poster	Thank you, colleagues. Let me clarify the question a little: is it possible to conclude that this system is self-learning? Or is it just that she is "pretrained"?

08-21-2021, 02:54 PM - 2 Likes	#6
vonBaloney Pentaxian Join Date: Oct 2011 Location: Albuquerque, NM Posts: 6,029	Originally posted by Grokh Thank you, colleagues. Let me clarify the question a little: is it possible to conclude that this system is self-learning? Or is it just that she is "pretrained"? 100% pretrained unless they say otherwise, which would be VERY surprising (computing power needed for training is immense). It could get updated models with firmware updates though (which is quite likely). I could even see them having a separate "model update" apart from the main firmware for that. But for a first version of this sort of thing there is no way they'd be jumping straight to self-learning. (Come back in 10 years for that maybe.)

08-21-2021, 02:59 PM	#7
Grokh Senior Member Join Date: Feb 2011 Location: St.-Petersburg Photos: Albums Posts: 267 Original Poster	Originally posted by vonBaloney 100% pretrained unless they say otherwise, which would be VERY surprising (computing power needed for training is immense). It could get updated models with firmware updates though (which is quite likely). I could even see them having a separate "model update" apart from the main firmware for that. But for a first version of this sort of thing there is no way they'd be jumping straight to self-learning. (Come back in 10 years for that maybe.) Thank you, I thought the processing power was too low, but suddenly ...

08-21-2021, 03:11 PM	#9
Kunzite Pentaxian Join Date: Apr 2007 Location: Romania Posts: 15,132	Originally posted by AfterPentax Mark II It is selflearning and that with the lapse of time that you take pictures (the more the better) with the K-3 Mark III, it recognizes the type of objects that you take pictures of. It is in this way that the predicative autofocus is trained to recognize your objects (read humans, birds, animals and so on) and it is used by the "predictive autofocus algorithms" as you put it. It means that the more you take pictures of birds or animals and other moving objects the better the respons of the autofocus with AF.C autofocus, because it learned to recognize the moving objects that you have taken pictures of. Hope this helps. I'm quite sure it doesn't learn as you use it. The "learning part" is very resource intensive, no camera should be capable of doing it. Possibly not even your PC. Then, I have doubts about the idea of training it to recognize (what?) just by using the camera, and no other clue. L.E. I see vonBaloney said something similar not long ago. Well, it's worth repeating.

08-21-2021, 03:28 PM - 1 Like	#10
AfterPentax Mark II Pentaxian Join Date: Mar 2021 Posts: 1,465	Originally posted by Kunzite I'm quite sure it doesn't learn as you use it. The "learning part" is very resource intensive, no camera should be capable of doing it. Possibly not even your PC. Then, I have doubts about the idea of training it to recognize (what?) just by using the camera, and no other clue. L.E. I see vonBaloney said something similar not long ago. Well, it's worth repeating. Yes it does, because it is a very simple system and the use of those words sets everybody on the wrong foot. It is maybe not "learning" but you can think of it as tagging things that share alike things. Recognizing eyes, can be a problem, because how does it know it is human or animal? They combine it with a laughing face which is mostly human. Eye and face recognition is wide spread in camera world. Microsoft Money to mention something a bit PC-like can recognize that you make the same kind of payments and then asks you if that is correct and if you want to tag it and each time such a payment is entered it will ask you if it must put an ok on it for this month. So if your camera recognizes something and focuses on that you can proceed or you can change the focus to the subject intended. You camera's programmed hardware can remember that. I edited my original post and elaborated on the subject with a few examples where it is used.

08-21-2021, 04:30 PM - 1 Like	#11
vonBaloney Pentaxian Join Date: Oct 2011 Location: Albuquerque, NM Posts: 6,029	If it uses the phrase "deep learning" that means neural networks -- that is a very specific thing and it is not a "general" phrase or marketing-speak. And I can assure you the camera is not doing neural network training. Again, maybe someday, but no way.

08-21-2021, 04:50 PM	#12
cfraz Veteran Member Join Date: Jan 2009 Location: Oregon, USA Posts: 521	Originally posted by AfterPentax Mark II Yes it does, because it is a very simple system and the use of those words sets everybody on the wrong foot. It is maybe not "learning" but you can think of it as tagging things that share alike things. Recognizing eyes, can be a problem, because how does it know it is human or animal? They combine it with a laughing face which is mostly human. Eye and face recognition is wide spread in camera world. Microsoft Money to mention something a bit PC-like can recognize that you make the same kind of payments and then asks you if that is correct and if you want to tag it and each time such a payment is entered it will ask you if it must put an ok on it for this month. So if your camera recognizes something and focuses on that you can proceed or you can change the focus to the subject intended. You camera's programmed hardware can remember that. I edited my original post and elaborated on the subject with a few examples where it is used. Absolutely not.

08-21-2021, 05:44 PM - 1 Like	#13
disconnekt Pentaxian Join Date: Oct 2011 Location: SoCal/I.E. Photos: Albums Posts: 1,701	Originally posted by vonBaloney 100% pretrained unless they say otherwise, which would be VERY surprising (computing power needed for training is immense). It could get updated models with firmware updates though (which is quite likely). I could even see them having a separate "model update" apart from the main firmware for that. But for a first version of this sort of thing there is no way they'd be jumping straight to self-learning. (Come back in 10 years for that maybe.) Im surprised Ricoh hasn come out & said "Hey Pentax users, share your photography photos with us so we can use the data we get from the photos to help improve the firmware we send out to make our DSLR's more accurate in various scenarios" Cuz as I see it, getting data directly from your user base on what they shoot (where exactly to improve on) will help out alot on witu future firmware updates/DSLR releases

08-21-2021, 07:03 PM	#14
jpzk Site Supporter Join Date: Jan 2009 Location: Québec Photos: Gallery \| Albums Posts: 8,251	Originally posted by AfterPentax Mark II It is selflearning and that with the lapse of time that you take pictures (the more the better) with the K-3 Mark III, it recognizes the type of objects that you take pictures of. It is in this way that the predicative autofocus is trained to recognize your objects (read humans, birds, animals and so on) and it is used by the "predictive autofocus algorithms" as you put it. It means that the more you take pictures of birds or animals and other moving objects the better the respons of the autofocus with AF.C autofocus, because it learned to recognize the moving objects that you have taken pictures of. Hope this helps. It is no marketing, but we all know that Japanese manufacturers (and many others) like to use difficult words to impress us, but it is really a simple learning process that is used by Honeywell, to mention a company, for adapting your central heating to your way of life and to weather conditions to start your central heating so that at a certain time, usually the time you wake up, the house is warm in the way you like it. This system of recognizing objects is also found in some photo processing programs that sort your pictures of your cats, your dogs your family in these specific categories if you want to. It also means that with the laps of time your camera will be responding faster. Deep learning, just sounds goods but what they meant is that it will take (a long) time . Would the K3III be able to "learn" while in AF.S mode rather than in AF.C and "extrapolate" this learning to AF.C ?

08-21-2021, 10:17 PM - 5 Likes	#15
kypfer Pentaxian Join Date: Jan 2018 Location: Jersey C.I. Posts: 3,595	Sorry, this concept that a camera could be "self-learning" has got to be mis-leading at the very least! If it were the case, within virtually no time at all every K-3iii would be "different" and in the case of needing repair, all this "learning" could be lost and the user would have to start over! Having to re-enter carefully crafted User Modes is pain enough (why oh why can't we save and load User Modes to/from SD card?), to have to "re-train" a camera that has made several (tens of) thousand exposures could be a major problem