Canon officially announces the Canon EOS-1D X Mark III

Nov 1, 2012
1,549
269
I think you will find the 1D X III learning has been tuned for the human head, as distinct from the human face. As I understand it this is to allow for the fast twisting and turning of the head in some sports and it allows for helmets, that is it still recognizes a helmeted head as the subject.

Maybe wildlife shooters will be able to buy firmware add-ons for specific animals!

I wish the deep learning could be taught, since my sports everybody wear funny outfits they are not recognized as human...
 
Upvote 0
Apr 25, 2011
2,510
1,885
Yes, that is true. It is possible to improve the learnt model and upload it via firmware. However, the way deep learning works is that each model learns one subject for max accuracy (e.g. a binary decision between face and not-face). The same model cannot be trained to learn face and bird and dog without compromising the accuracy.
Actually, co-training of a neural network for multiple goals (outputs) usually improves its generalization abilities, reduces overfitting on the training set and improves the network's accuracy during real-world use.

At least as believed by the industry right now.
 
  • Like
Reactions: 1 user
Upvote 0
Hard to get a real sense from all the comments if contributors here are super excited at the 1DX III that's been produced or whether its more of a damp squid.
It overall just doesn't seemed to be designed for serious amateur photographers.

'why 20 million pixels'
"Canon consulted heavily with major press and photo agencies ... "

It really does seem to be designed for press agencies to able to photograph and video with the same piece of gear.
I personally don't get why a press agencies would declare a 20MP limitation as a key requirement.
20 / 22/ 24 wouldn't make a huge difference, the speed of telecommunictions is only getting quicker.
No issue with transmitting 4K video, only with 24mp photos.

The bulk of Canon 1dx sales must come from press and photo agencies.
I think a modest MP increase would have helped increase the number of upgrades from 1DX II.
I was going to upgrade but I am going to hold off for the moment.
I'll follow the reviews and see if the focusing system is that good that it justifies an upgrade.
7599 Euros is a high price to pay for an upgrade that has few spec improvements.
It's all down to its focusing ability and how it handles in expert users hands.

I am pretty excited. This camera is a powerhouse and I am buying one. If it had 30mp I would be sleeping in a tent in front of Canon trying to be first in line to get it. As it is I am preordering from the comfort of my office via the internet.
 
Upvote 0

Travel_Photographer

Travel, Landscape, Architecture
Aug 30, 2019
94
126
Having said I won't buy, it's the AF that keeps catching my attention. https://www.canon-europe.com/cameras/eos-1d-x-mark-iii/autofocus/ I have spent many hours trying to catch small birds in great action poses and in spite of fiddling with the case settings never got what I would describe as true auto focus that could catch scenes like the one I'm posting. If that's possible now ... oh dear ... lets see, how can I scare up the cash.View attachment 188139View attachment 188140View attachment 188141


I bought the R and quickly convinced my daughter it was perfect for her. Having filter capability with the R and 11-24 was a big deal so yes I'm thinking R but it would have to be 1 level or I'll just never quit griping. The video on the R was so poor compared to the 1DX2, especially the jello. You have to shoot 1 level to appreciate the finer points it offers. I remain conflicted.

BTW, I believe the shots I posted were 14 FPS in case someone is inclined to think 5 FPS would do the job. Is it just too much to expect a bird to be focused in this scenario? I think the only way is deep learning with bird photos.

Jack

Right, what I'm suggesting, since autofocus seems to be one of your primary considerations, is that you wait for the pro-level R mount camera that many people are expecting to be announced (hopefully) sometime soon.

The 1-series is excellent for those that need an optical viewfinder for fast moving subjects where EVF lag might post a problem, for top build quality and weatherproofing, high FPS, etc. That said, it is my opinion (any many others) that dual-pixel autofocus on the sensor is far superior to the focus mechanism used when shooting through an optical viewfinder. That's in large part why so many are so highly anticipating a pro-level R.

Regarding your hummingbird, I'm not sure what high FPS has to do with your focus issues. You're correct in that those aren't particularly in focus, but high FPS is not your issue. What shutter speed are using? Yes, I know the hummingbird's wings flap really fast, I'm not suggesting that you used a slow shutterspeed because the wings are blurry, it just seems the whole images are fairly blurry so I'm thinking maybe an even faster shutter speed might have helped? And perhaps the aperture was too wide and you moved backward or forward at the time you pressed the shutter? Were you in Servo mode? If you weren't in servo and you moved even a fraction of an inch after autofocusing, the whole image would be blurry.

Either way, obviously people, including me, have been successfully focusing on and getting tack-sharp images of hummingbirds and other fast-moving birds before these "deep learning algorithms" came around. No algorithm is going to make up for incorrect camera settings to begin with (I'm not suggesting you used incorrect settings, just making a general statement). I have no idea why those hummingbirds are out of focus, but I can assure you it's not a limitation of the camera. Personally, I think you'd love a pro-level R mount.
 
Upvote 0

Jack Douglas

CR for the Humour
Apr 10, 2013
6,980
2,602
Alberta, Canada
Right, what I'm suggesting, since autofocus seems to be one of your primary considerations, is that you wait for the pro-level R mount camera that many people are expecting to be announced (hopefully) sometime soon.

The 1-series is excellent for those that need an optical viewfinder for fast moving subjects where EVF lag might post a problem, for top build quality and weatherproofing, high FPS, etc. That said, it is my opinion (any many others) that dual-pixel autofocus on the sensor is far superior to the focus mechanism used when shooting through an optical viewfinder. That's in large part why so many are so highly anticipating a pro-level R.

Regarding your hummingbird, I'm not sure what high FPS has to do with your focus issues. You're correct in that those aren't particularly in focus, but high FPS is not your issue. What shutter speed are using? Yes, I know the hummingbird's wings flap really fast, I'm not suggesting that you used a slow shutterspeed because the wings are blurry, it just seems the whole images are fairly blurry so I'm thinking maybe an even faster shutter speed might have helped? And perhaps the aperture was too wide and you moved backward or forward at the time you pressed the shutter? Were you in Servo mode? If you weren't in servo and you moved even a fraction of an inch after autofocusing, the whole image would be blurry.

Either way, obviously people, including me, have been successfully focusing on and getting tack-sharp images of hummingbirds and other fast-moving birds before these "deep learning algorithms" came around. No algorithm is going to make up for incorrect camera settings to begin with (I'm not suggesting you used incorrect settings, just making a general statement). I have no idea why those hummingbirds are out of focus, but I can assure you it's not a limitation of the camera. Personally, I think you'd love a pro-level R mount.

This was just a sequence that popped up as I searched quickly for something to illustrate. I am fully aware of the human problem (me) but less aware of how to solve it - the teach old dogs (me) new tricks issue. :unsure: Your points are all valid. :)

Shooting hummingbirds in Costa Rica was one of the easiest tasks as far as getting the birds in focus and I'm not talking feeders, unless you consider flowers to be feeders, but having a frame that would keep them in focus as they dart around would be amazing with the high FPS. Here is another example of where fast AF could be helpful and where high FPS also is very helpful as far as pose is concerned. It doesn't help to often be shooting 400 X2 and with poor lighting, and that's the dilemma, there are always trade-offs.

Jack

2 Death to You_s_30889.JPG
 
  • Like
Reactions: 2 users
Upvote 0
Could you please elaborate on what you mean here? Neural Networks can be used for classification problems even if there are more than two cases in the output, so I'm not sure what you mean here.
I think you train a neural net model with positive examples (e.g. human head) and negative examples (e.g. cat, dog, flower, cloud, etc.) of roughly equal set size, say 100k pictures in each set, and then when using the trained model in the camera, when seeing a scene it can decide whether it is a human head or not (binary decision) with high accuracy. Of course it is possible to train a model for more objects using the same principle.
 
  • Like
Reactions: 1 user
Upvote 0

Travel_Photographer

Travel, Landscape, Architecture
Aug 30, 2019
94
126
This was just a sequence that popped up as I searched quickly for something to illustrate. I am fully aware of the human problem (me) but less aware of how to solve it - the teach old dogs (me) new tricks issue. :unsure: Your points are all valid. :)

Shooting hummingbirds in Costa Rica was one of the easiest tasks as far as getting the birds in focus and I'm not talking feeders, unless you consider flowers to be feeders, but having a frame that would keep them in focus as they dart around would be amazing with the high FPS. Here is another example of where fast AF could be helpful and where high FPS also is very helpful as far as pose is concerned. It doesn't help to often be shooting 400 X2 and with poor lighting, and that's the dilemma, there are always trade-offs.

Jack

View attachment 188157

Beautiful photo. Really spectacular!

Yep, understood. I was just tossing another alternative into the mix with mirrorless, since you were seeking some input. I was in the same boat regarding AF challenges as I was trying to figure out the best way to focus on the wildest creature of them all: Toddlerus Runnerus (Common Name: Running Toddler). I tried an RP and was pleasantly shocked at how effortlessly and intuitively it focuses. While looking through the EVF, I use my thumb on the touchscreen to move the AF box over whatever moving subject I'm trying to focus on, and that's it. It uses over 4000 AF points on the sensor to instantly track the subject wheverever it moves over almost the entire frame. Anyway, I'm not trying to convince you, just sharing my own experiences! If a pro R body comes out with high FPS and weather sealing, I will likely snap it up in a second.
 
  • Like
Reactions: 1 user
Upvote 0
I was torn between going for higher resolution via 5D4 or 1 level camera via 1DX2. It went on for weeks and finally my wife nudged me over the edge - she sees 1D as "the best". Essentially it was a coin toss. To compensate for no bump in MPs (sold the 6D) I bought the 400 DO II, really over-spending and rationalized that 400 X2 based on reputable comments would be very satisfactory. Truth is, 800 did put me within reach for a lot of birding situations but as more samples of the results of high MP shots passed by me I came to realize that my cropping capabilities were now more limited if I wished to maintain the detail and sharpness I was accustomed to (I had previously used the 300 2.8 II, which uncropped was amazing).

As a retired person and just an amateur photographer what in the world is wrong with me not just being satisfied with what I've got?! Well, the problem is I tend towards being a perfectionist; nothing to do with showing off, I just love excellent results. Honestly, I can't afford this hobby but by cutting out other likes, I do, and I really enjoy it.

With over 3 years of fun with the 1DX2, I've absolutely loved the 14 FPS and the 4K60 and could have been quite happy with a boost of 4 or slightly more MPs and in that case jumped on the expensive 1DX3 bandwagon. Others say no, but I maintain that often I'm on the edge, where cropping that would be offset by those extra MPs, would be very much appreciated. Perhaps by analogy it could be compared to having a car that tops out at 55 when speed limits are 60.

On the other hand, talking about being on the edge, recently in Costa Rica, I was reminded what it was like to have roughly 12 hours of light and dark and mist and deep shadows in jungles and, well, there is no doubt that low noise higher ISO would have been really helpful especially when I was at F4 X2. In many of those cases my reach with 800 (560 probably would have worked too) was fine but the lighting just wasn't there.

So, which poison do I choose. What I'm reading about the 1DX3 resonates because I am very aware of improvements that could be game changing if the reports are trustworthy but would I ever be happy with just 20 MPs. Anyone else feel like me, or share the same thoughts? ;) :cry::unsure::)

Jack
Is one option to keep what you have, and spend the difference between trading in your 1DX II and purchasing a 1DX III ($3-4k) instead on the forthcoming high MP R? - that way you have the best of both worlds? (albeit, spread over two cameras!)
 
  • Like
Reactions: 1 user
Upvote 0

Joules

doom
CR Pro
Jul 16, 2017
1,801
2,247
Hamburg, Germany
I think you train a neural net model with positive examples (e.g. human head) and negative examples (e.g. cat, dog, flower, cloud, etc.) of roughly equal set size, say 100k pictures in each set, and then when using the trained model in the camera, when seeing a scene it can decide whether it is a human head or not (binary decision) with high accuracy.
There are a lot of different neural network architectures and they all have their own ways of working. But you are definitely not constrained to a binary decision. Neural networks are just functions, you apply them to an input and get an output. The output can be of any dimension though, it doesn't have to be a two dimensional "Yes" or "No" output. With your example, you can have an output with 4 numbers, where the highest one can be interpreted as the recognised result. So if your first number is for cat, second for dog, third for flower and fourth for cloud, and your numbers are 0.2 0.98 0.1 and 0.0 you should interpret that as "Your input corresponds to a dog".
 
  • Like
Reactions: 1 users
Upvote 0
Jul 28, 2015
3,368
570
I was torn between going for higher resolution via 5D4 or 1 level camera via 1DX2. It went on for weeks and finally my wife nudged me over the edge - she sees 1D as "the best". Essentially it was a coin toss. To compensate for no bump in MPs (sold the 6D) I bought the 400 DO II, really over-spending and rationalized that 400 X2 based on reputable comments would be very satisfactory. Truth is, 800 did put me within reach for a lot of birding situations but as more samples of the results of high MP shots passed by me I came to realize that my cropping capabilities were now more limited if I wished to maintain the detail and sharpness I was accustomed to (I had previously used the 300 2.8 II, which uncropped was amazing).

As a retired person and just an amateur photographer what in the world is wrong with me not just being satisfied with what I've got?! Well, the problem is I tend towards being a perfectionist; nothing to do with showing off, I just love excellent results. Honestly, I can't afford this hobby but by cutting out other likes, I do, and I really enjoy it.

With over 3 years of fun with the 1DX2, I've absolutely loved the 14 FPS and the 4K60 and could have been quite happy with a boost of 4 or slightly more MPs and in that case jumped on the expensive 1DX3 bandwagon. Others say no, but I maintain that often I'm on the edge, where cropping that would be offset by those extra MPs, would be very much appreciated. Perhaps by analogy it could be compared to having a car that tops out at 55 when speed limits are 60.

On the other hand, talking about being on the edge, recently in Costa Rica, I was reminded what it was like to have roughly 12 hours of light and dark and mist and deep shadows in jungles and, well, there is no doubt that low noise higher ISO would have been really helpful especially when I was at F4 X2. In many of those cases my reach with 800 (560 probably would have worked too) was fine but the lighting just wasn't there.

So, which poison do I choose. What I'm reading about the 1DX3 resonates because I am very aware of improvements that could be game changing if the reports are trustworthy but would I ever be happy with just 20 MPs. Anyone else feel like me, or share the same thoughts? ;) :cry::unsure::)

Jack

I'm with you every step of the way, there Jack. The main reason I bought the 1Dx2 was that the 7D2 AF can be a bit twitchy and because most of my good wildlife shots are taken in morning or late afternoon the 7D2 is often pushing it, especially for BIF. I have also found that the 7D2 starts to struggle with the 2x converter on the 400DOii, which means that for BIF 7D2+400+1.4x has about the same sort of reach as 1Dx2+400+2x.

Also, largely because of the cropping, I have learned that I don't need to have frame-filling images every time (which has actually benefitted my composition and editing). Having said all that, I do find the sheer quality of pixels on the 1Dx2 means even with the same lens on both cameras I can crop almost to the same FOV as the 7D2 with very little loss in picture quality.
BUT....give me a 5DIV with a bit extra battery oomph and 8-10 fps and I would be well happy.

Why oh, why do I like the challenge of BIF so much. Why can't I just be happy with a bird on a stick?
 
  • Love
Reactions: 1 user
Upvote 0
Jul 28, 2015
3,368
570
No. The deep learning is learnt and burnt in. It doesn’t learn anything new but has enough stored data to estimate the true subject and maintain the af on what it has already been told is probably the subject. But this is only in full area mode, you can use the camera as a MkII in the other af modes and bypass the ‘learning’ which truly is a misnomer.

If that is the case, if it is baked in, then as you say it is not 'learning' - it is simply using a more detailed algorithm which would mean that Canon is using modern buzzwords to sell it stull (like 'nano' was few years ago, and not is it 'i-whatever').
When the Olympus E-M1X came out, there was talk that it actually 'learnt' the photographer's preferred subjects. So with a basic algorithm, it knows what a head looks like but if the head was turned sideways it initially would miss it. But gradually over time it would see a head turned a bit, then turned more, then turned more and as the photographer shot more and more heads it would 'learn'. Similarly for cars etc. I don't know how true any of that is, but it strikes me as a method with more potential for tracking the subject across the frame or in front of complex backgrounds.
 
  • Like
Reactions: 1 user
Upvote 0

unfocused

Photos/Photo Book Reviews: www.thecuriouseye.com
Jul 20, 2010
7,184
5,483
70
Springfield, IL
www.thecuriouseye.com
If that is the case, if it is baked in, then as you say it is not 'learning' - it is simply using a more detailed algorithm which would mean that Canon is using modern buzzwords to sell it stull (like 'nano' was few years ago, and not is it 'i-whatever')...

Always helps to go to the source. Canon White Paper:

Head Detect AF is supported by “deep learning.” Canon engineers have established a database of thousands of images of heads in varying types of positions, including turned away from the camera, and even wearing various types of sports helmets. Deep learning has been applied to the AF system (within the DIGIC X processor) to leverage this info for immediate recognition of human subjects, when traditional Face Detect AF cannot read sufficient information to confirm a subject.

Interestingly, this transition from Face Detect AF to Head Detect AF leverages both the DIGIC 8 processor (normally the primary source of AF tracking control) and the more powerful, central DIGIC X processor. DIGIC X is the source of Head Detect AF data, and its deep learning support library — but the transition from Face Detect AF to Head Detect AF is extremely smooth, even in the middle of a 16 fps shooting sequence.

This does make it possible that future firmware updates would add more details/information to the camera's "deep learning" database.
 
Upvote 0

koenkooi

CR Pro
Feb 25, 2015
3,574
4,109
The Netherlands
Always helps to go to the source. Canon White Paper:
[..]


This does make it possible that future firmware updates would add more details/information to the camera's "deep learning" database.

Or make it possible to replace it with something else! So where's the place on the Canon website I can upload all my jumping spider pics to?
 
Upvote 0

Russ6357

CR Pro
Dec 17, 2019
18
20
As always it’s not the buzzword tech that’s the important thing but the quality of implementation- how large and well labeled the data set, what model used and so forth.
As noted the output need not be Y:N but instead, for example, a series of probabilities, e.g.

(Is a focus target) (Is human) (Is bird) (Is dog) (Is cat)

0.98, 0.03, 0.11, 0.65, 0.89

which would relate to start focusing with an focus priority assuming Cat but it’s a bit dog like too...
 
Upvote 0

Jack Douglas

CR for the Humour
Apr 10, 2013
6,980
2,602
Alberta, Canada
Pre Ordered...

What a sucker!

OK, since you called yourself a sucker, who am I to challenge that. :)

If I were not financially constrained I wouldn't blink before ordering it. Knowing Canon and knowing Canon's history, there is not a shadow of doubt in my mind that the shortcomings of the Mark II have been largely dealt with and the camera will be a significant improvement in the areas that Canon has felt pressured to improve (AF was one of them) and where they were able to improve (speed was obviously a limiting factor and that's been drastically improved via CFexpress, Digic X etc.).

I've always shook my head when people praised the joystick because it clearly was clumsy when speedy AF point movements were required; now the new AF-ON button appears to be a godsend. I also have no doubt that the AF capability of the latest Nikon and Sony top cameras has exceeded what we got with the Mark II, even considering that many users did not adjust the parameters as optimally as possible. I suspect that will no longer be the case, at least for the time being.

In the realm of video, those in the know are expressing overall satisfaction with the added capability and already the quality of the Mark II was generally praised except for file size. The dumb CF, CFast arrangement is gone and for those with deep pockets the new cards offer many advantages.

My concern, if I were to purchase, is in the realm of, I can only justify so much expenditure on a strictly luxury item when there are other needs and wants and I already have a very good camera. My "worry" is that there may be significant improvements relative to a 1 level offering in the R line and if that camera will be slightly smaller and significantly lighter then as an aging user I'll regret that my (probably) last camera is a beast.

I seldom express frustration with Canon's approach to making money but if I were to express one major gripe it would be that there needs to be two somewhat different 1 level cameras, one that maximizes speed and one that is still fast but optimizes resolution for those of us that invariably do more cropping. I don't get the logic of forcing those with resolution needs into a lower than 1 level feature set.:unsure:

Jack
 
  • Like
Reactions: 6 users
Upvote 0