Patent: Microadjustment Automated

[jrista]

I'd have to read through the patent to see if there are any unique elements to it but generally speaking the idea of using CDAF to auto-tune PDAF has lots of prior art online.

It really depends on the implementation. Use of CDAF to tune PDAF has been done, but patents are more about the specifics of the implementation as well as the description of design that employ a certain practice. Overly broad and vague patents, like Apples bogus claim to patent of "rounded corner icons", are bound to fail due to prior art. If Canon has a very specific, technical implementation that relies on some very specific aspect of their technology, such as DPAF, then it is possible it would not be rendered invalid by prior art.

 

[Don Haines]

I'd have to read through the patent to see if there are any unique elements to it but generally speaking the idea of using CDAF to auto-tune PDAF has lots of prior art online.

It really depends on the implementation. Use of CDAF to tune PDAF has been done, but patents are more about the specifics of the implementation as well as the description of design that employ a certain practice. Overly broad and vague patents, like Apples bogus claim to patent of "rounded corner icons", are bound to fail due to prior art. If Canon has a very specific, technical implementation that relies on some very specific aspect of their technology, such as DPAF, then it is possible it would not be rendered invalid by prior art.

Exactly!

The concept might be public use, but a specific algorithm could be patented... In fact, they would want to patent it to prevent someone else from copying their algorithm, patenting it, and suing Canon...

 

[hpjfromdk]

For those interested, the English version of JP2013254194, namely patent app US20120300116 was issued back in Nov 29 2012, can be found here:

http://www.freepatentsonline.com/20120300116.pdf

It's probably a little easier to read than the JPO machine translations..

 

[AvTvM]

MILCs only need one or at most 2 AF-systems ... contrast-detect [speed is solely dependent on porocessing power and algorithms available] and/or on-sensor phase-detect.

Not quite. One big difference between phase detect AF (PDAF) and contrast detect AF (CDAF) is that the phase difference provides data on both the magnitude of the change needed to achieve focus and the direction of the change, whereas CDAF does not provide directional information. So, for CDAF to work, it has to pick a direction and start to move the focusing elements, then determine whether the image becomes more or less focused - there's a 50/50 chance of getting it right (perhaps higher if the system guesses based on where the focusing group starts, e.g. if near the MFD, guess that focus needs to move away not toward, but it's still a guess). A big boost in speed when supplementing CDAF with on-sensor PDAF is that the focus starts moving in the right direction first, every time.

Yes, i agree on the advantages of combining on-sensor pdaf (=no mirrors onvolved) plus cd-af .. As this hybrid combo combines the relative strengths of both concepts.

But ... CD-AF finds its target without pdaf as well. As proven in gazillions of smartphones and digital cameras of all sorts. And if there is enough cpu power snd smart algorithms it cokld be as fast as you want it. 100 fps full tracking at any point of the imaging surface. Thats why i believe that at the end of the day we will habe solid state cameras with cd-af only. Mechanics and pdaf - even the on-sensor variations - will disappear.

Until then i would be happy to get a 36x24 sensored camera with only 2 instead of 3 separate af systems.

 

[Loswr]

But... CD-AF finds its target without pdaf as well. As proven in gazillions of smartphones and digital cameras of all sorts. And if there is enough cpu power snd smart algorithms it cokld be as fast as you want it.

Yes, gazillions of smartphones and digital cameras that focus very slowly. CDAF might be "fast enough" but PDAF or a hybrid system will always be just a little bit faster, because the direction of movement is known at the outset.

 

[Albi86]

But... CD-AF finds its target without pdaf as well. As proven in gazillions of smartphones and digital cameras of all sorts. And if there is enough cpu power snd smart algorithms it cokld be as fast as you want it.

Yes, gazillions of smartphones and digital cameras that focus very slowly. CDAF might be "fast enough" but PDAF or a hybrid system will always be just a little bit faster, because the direction of movement is known at the outset.

I doubt the day is far away when such difference will become field irrelevant, however relatively big it might be.

One point, however, that maybe hasn't been considered yet: what about the lenses? The problem, if I understand correctly, might arise with primes that don't have high contrast wide open. We might need to get to a point where the average lens performs wide open like the latest Zeiss or Sigma to make contrast AF reliable.

 

[jrista]

MILCs only need one or at most 2 AF-systems ... contrast-detect [speed is solely dependent on porocessing power and algorithms available] and/or on-sensor phase-detect.

Not quite. One big difference between phase detect AF (PDAF) and contrast detect AF (CDAF) is that the phase difference provides data on both the magnitude of the change needed to achieve focus and the direction of the change, whereas CDAF does not provide directional information. So, for CDAF to work, it has to pick a direction and start to move the focusing elements, then determine whether the image becomes more or less focused - there's a 50/50 chance of getting it right (perhaps higher if the system guesses based on where the focusing group starts, e.g. if near the MFD, guess that focus needs to move away not toward, but it's still a guess). A big boost in speed when supplementing CDAF with on-sensor PDAF is that the focus starts moving in the right direction first, every time.

I don't think CDAF's lack of predetermine directionality would matter in this case. All that is necessary is for CDAF to achieve focus as a reference point. Once a reference point is attained (and CDAF CAN indeed achieve very good focus once it's done going through all its gyrations), you save the focus group position of the lens, then all you need to do is test PDAF at a distribution of AFMA settings until you zero in on the one that most closely matches the CDAF position, and ensure that setting produces repeatable results.

 

[horshack]

I don't think CDAF's lack of predetermine directionality would matter in this case. All that is necessary is for CDAF to achieve focus as a reference point. Once a reference point is attained (and CDAF CAN indeed achieve very good focus once it's done going through all its gyrations), you save the focus group position of the lens, then all you need to do is test PDAF at a distribution of AFMA settings until you zero in on the one that most closely matches the CDAF position, and ensure that setting produces repeatable results.

The PDAF cycles aren't required. Once CDAF has found the optimum contrast point, the AFMA value can be determined almost instantly by evaluating the phase differential and then applying the appropriate AFMA adjust value to bring that differential to its minimum. This is one method outlined in Canon's patent.

 

[Loswr]

MILCs only need one or at most 2 AF-systems ... contrast-detect [speed is solely dependent on porocessing power and algorithms available] and/or on-sensor phase-detect.

Not quite. One big difference between phase detect AF (PDAF) and contrast detect AF (CDAF) is that the phase difference provides data on both the magnitude of the change needed to achieve focus and the direction of the change, whereas CDAF does not provide directional information. So, for CDAF to work, it has to pick a direction and start to move the focusing elements, then determine whether the image becomes more or less focused - there's a 50/50 chance of getting it right (perhaps higher if the system guesses based on where the focusing group starts, e.g. if near the MFD, guess that focus needs to move away not toward, but it's still a guess). A big boost in speed when supplementing CDAF with on-sensor PDAF is that the focus starts moving in the right direction first, every time.

I don't think CDAF's lack of predetermine directionality would matter in this case. All that is necessary is for CDAF to achieve focus as a reference point. Once a reference point is attained (and CDAF CAN indeed achieve very good focus once it's done going through all its gyrations), you save the focus group position of the lens, then all you need to do is test PDAF at a distribution of AFMA settings until you zero in on the one that most closely matches the CDAF position, and ensure that setting produces repeatable results.

No, it doesn't matter for AFMA. But routine focusing during everyday shooting was the subject under discussion (albeit one tangential to the topic).

 

[jrista]

I don't think CDAF's lack of predetermine directionality would matter in this case. All that is necessary is for CDAF to achieve focus as a reference point. Once a reference point is attained (and CDAF CAN indeed achieve very good focus once it's done going through all its gyrations), you save the focus group position of the lens, then all you need to do is test PDAF at a distribution of AFMA settings until you zero in on the one that most closely matches the CDAF position, and ensure that setting produces repeatable results.

The PDAF cycles aren't required. Once CDAF has found the optimum contrast point, the AFMA value can be determined almost instantly by evaluating the phase differential and then applying the appropriate AFMA adjust value to bring that differential to its minimum. This is one method outlined in Canon's patent.

Well there you go then. Even better. Wasn't there a thread recently about the lack of Canon innovation? I think this patent would debunk that notion right off.

 

[scottkinfw]

Exactly (and I'm not a golfer)

So instead of trying to fix the AF issues, they come up with a ridiculous AFMA solution? At least for me, the AF accuracy and AFMA value depend heavily on the distance to the subject and even the lighting can affect results. Having 1 value for a lens does not fix much. Besides, a slicing golfer can surely aim off fairway at the tee, but that is a lame way to deal with the slice. A better solution is to fix the slice.

 

[Loswr]

Exactly (and I'm not a golfer)

So instead of trying to fix the AF issues, they come up with a ridiculous AFMA solution? At least for me, the AF accuracy and AFMA value depend heavily on the distance to the subject and even the lighting can affect results. Having 1 value for a lens does not fix much. Besides, a slicing golfer can surely aim off fairway at the tee, but that is a lame way to deal with the slice. A better solution is to fix the slice.

I'm sure Canon could 'fix the slice' i.e. tighten manufacturing tolerances sufficiently to obviate the need for AFMA. But if they did, you'd need Tiger Woods' income to buy those lenses.

Many other 'issues' that plague lenses, like focus breathing, repeatability of setting a focus distance, etc., are correctable. Canon has done so for their Cinema lenses (which don't even have AF) - check the price tag of those when wishing for Canon to 'fix the AF issues.'

 

[AvTvM]

Well there you go then. Even better. Wasn't there a thread recently about the lack of Canon innovation? I think this patent would debunk that notion right off.

Well ... n automated in-camera AF-calibration system certainly is no huge "innovation".
Basically it is just a necessary little service fix to avoid problems stemming from manufacturing tolerances in camera and lens assemblies. The very idea to include a storage for AF-system correction factors in a camera [AFMA] and to automaticallycalibrate the phase-AF system with a second AF-system present in all current Canon digital cameras, is not very innovative per se. Even more so, when out of camera software to this purpose already has been invented. ANd notwithstanding the engineering finesse it may take to implement in camera.

Anway, AFMA is a solution to an "intermediate problem" that will automatically disappear, as soon as mirrors in cameras' lightpath and off-sensor (phase) AF-systems will be abolished altogether.

I'd rather Canon to focus their innovative powers towards a really good, FAST AND PRECISE FOCUSING, MOTION-TRACKING-CAPABLE AF-system in a 35-sensored solid-state camera. 8)

 

[Zv]

Does this mean no more trying to line up targets? Or is it just that the software to determine if the lens is out of whack is now installed in the cameras firmware as opposed to your laptop? And would it set just one value or more?

Would be awesome if it could learn and adjust as you shoot.

 

[madmailman]

Would the next "logical" step be using CDAF on one of the focus points and then PDAF takes over in AI-Servo (or video) mode to enable focus anywhere on the sensor (with a STM lens no doubt)? A combo AF system for AI-servo mode sounds like it could be very useful to a lot of people. No?

 

[rs]

Does this mean no more trying to line up targets? Or is it just that the software to determine if the lens is out of whack is now installed in the cameras firmware as opposed to your laptop? And would it set just one value or more?

Would be awesome if it could learn and adjust as you shoot.

Maybe I'm missing something, but I'd have thought that as there's an inevitable time delay between AF chip based PDAF and sensor based CDAF (the mirror has to move between these two states), the only sensible way to accomplish a sync/calibration between the two is to have the camera fixed and pointing at a suitable target. So I'm guessing it'll just be a glorified in-camera version of plugging your laptop in and paying for suitable software.

Learning as you shoot could be the next logical step, but with some potential issues. The camera knows which AF point(s) it used, and which one locked on the best, so it knows which part of the image should have the greatest micro contrast. A simple analysis of the photos taken could be used to provide feedback of how well it performs, and make a best guess at how much to tweak, even if it doesn't know which direction to tweak it in (although if it records how far out each PDAF point was, it could use a best fit of contrast in all those areas to speed things up). After a handful of photos, a minor AF issue could be completely resolved. Over time it could also be used to learn different zoom settings, apertures and focus distances this way. If it's out by more than a certain tolerance, it would take too long and cause too many misfocused shots - not great if you depend on the camera/lens to deliver the goods there and then. In that case it could suggest you sort it out with a full on calibration as above. It could also learn on the fly about the way the photographers subjects move to predict AF better, and to learn about the photographers technique. The focus and recompose technique will completely throw that system though...

 

[rs]

Would the next "logical" step be using CDAF on one of the focus points and then PDAF takes over in AI-Servo (or video) mode to enable focus anywhere on the sensor (with a STM lens no doubt)? A combo AF system for AI-servo mode sounds like it could be very useful to a lot of people. No?

If you're using the optical viewfinder for stills, sensor based CDAF can't be used as the mirror blocks all light to the sensor except during the very brief moment the photo is taken. An AF chip based CDAF would suffer from the same misalignment issues as AF chip based PDAF, so no gains in doing it that way...

 

[epsiloneri]

Hm, this sounds very much like what I suggested in this forum, 3 years ago:

1) AF focus on a target, say a properly aligned focus target. The camera registers what it thinks is the best focus.
2) Without moving the camera or the focus tagert, go to 10x live view and manually focus to what you think is the best focus, push a button or something for the camera to register what your preferred focus is.
3) The camera makes use of info from 1 and 2 to compute MA.
[...]
Alternatively, one could let the camera itself compare the AF between the AF sensors and the live view contrast AF, and compute MA under the assumption that live view AF is more accurate. That would be even simpler, and according to my experience, live view contrast AF is nearly always accurate (but slower). Contrast AF is not affected by front/back focus issues, since it uses the actual detected image for AF, so it would be perfect to correct for AF sensor MA. I can imagine setting up the camera on a tripod and align it to a focus target, select "calibrate AF" from a camera menu, and then let the camera automatically cycle through 10 AF measurement cycles (say), computing the best MA. Why not, Canon?

I'm glad Canon listened I hope this makes it into firmware at some point...

 

[Marsu42]

Automation of the AF adjustment

Magic Lantern dot_tune module does this, cost: zero, availability: now ... there may be concerns over precisions and whatnow, but I'm positive it's quick & worth a shot, esp. over manual attempts. Works for me.

 

[TWI by Dustin Abbott]

Automation of the AF adjustment

Magic Lantern dot_tune module does this, cost: zero, availability: now ... there may be concerns over precisions and whatnow, but I'm positive it's quick & worth a shot, esp. over manual attempts. Works for me.

Let's hope this isn't the last of the Magic Lantern advances that makes it into official Canon firmware. Canon should be thankful for the ML team, as they give not only some great advances/flexibility to Canon cameras, they also give the "tuning/customizing" crowd something real and tangible to do to tweak their cameras in a non-destructive way.