Pana 20mm F1.7 focus speed vs Oly 45mm F1.8 (video)

Started Mar 9, 2013 | Discussions thread
Anders W
Forum ProPosts: 19,064Gear list
Like?
Re: Pana 20mm F1.7 focus speed vs Oly 45mm F1.8 (video)
In reply to micksh6, Mar 17, 2013

micksh6 wrote:

Anders W wrote:

micksh6 wrote:

Anders W wrote:

micksh6 wrote:

Anders W wrote:

micksh6 wrote:

Anders W wrote:

micksh6 wrote:

Anders W wrote:

micksh6 wrote:

Anders W wrote:

micksh6 wrote:

Anders W wrote:

micksh6 wrote:

Anders W wrote:

micksh6 wrote:

I saw your test and I wanted to compare lenses that I also have because I find the following results from your test unusual:

Panasonic 20/1.7
Test 1: 0.33 s
Test 2: 0.48 s
Olympus 45/1.8
Test 1: 0.39 s
Test 2: 0.46 s

Anders, on one hand you admit that 20mm focuses slower because when it hunts it hunts longer. On the other hand, your results show very similar focusing speed. How would you explain that?

Very easy. I have mentioned it many times before. As long as you refocus on a target reasonably close to the previous one (where going from infinity to 75 cm as in my test defines the outer limit of "reasonably close"; note here that refocusing from 75 to 20 cm is outside the limit since the amount the lens elements have to travel is greater when going from 75 to 20 cm than when going from infinity to 75 cm) the impact of the slower max barrel movement of the 20/1.7 will be marginal. The AF time in this case is primarily a matter of body processing time. When the lens hunts from one end of the focus range to the other (infinity to 20 cm and back again), the slower max barrel movement of the 20/1.7 in comparison with other lenses (with internal focus) will start to make a difference.

This explanation isn't good enough for me. Regardless of how many focusing steps are done the proportion between body processing time and motor movement time should stay the same. For each motor step there is one processing step. The focusing time rather depends on number of steps, and not on total distance. That's why 4/3 DSLR lenses not optimized for CDAF focus so slow on m4/3. I hope you won't argue at least with that.

Could you please elaborate on the reasons why you think this theory of yours explains why FT lenses designed for PDAF are so much slower to AF on MFT cameras than are native MFT lenses. I think I already know why, but I am not sure, based on what you say above, that you do.

With PDAF lens focusing element moves to focus position in one step (maybe 2-3 if camera does final focus adjustment).
With CDAF the focusing element makes dozens of steps before it settles in focus position.

No. CDAF doesn't move in steps. See here for the way it actually works:

http://forums.dpreview.com/forums/post/51012091

Yes, the way I described is precisely how CDAF works. Move lens, stop lens, capture image sample, measure contrast, move lens again, stop lens, capture another sample, and so on. Find the best contrast after multiple measurements. Ability to start and stop lens focus element fast is the crucial thing for speed.

There is no need to stop the lens in order to read the sensor or measure contrast. The movement of the focus mechanism, the reading of the sensor, and the evaluation of contrast can all take place simultaneously. We can easily observe this with our MFT equipment. There is no studder in the movement as there is with PDAF.

The lens absolutely needs to fully stop when camera is taking image sample for contrast measurement. Are you seriously thinking that image samples are taken while the lens is in motion? No, it can't work that way - there would be blur because of the motion.

Yes, I am seriously thinking precisely that. Each frame will simply reflect the average focus point used during the time of capture. No blur is added by the AF movement. If the AF mechanism is moving in the right direction, the average focus point used for the current frame will yield higher contrast than that used for the previous one.

With the sequential scenario you outline (start lens AF movement, stop lens AF movement, expose sensor, read sensor, evaluate data, start lens AF movement again, and so on) there would be no chance of reaching the short AF times we currently enjoy. These are made possible by having these processes occur in parallel rather than sequentially. Lens AF movement, sensor exposure, sensor read-out, and data evaluation occur in parallel and continuously.

We enjoy pixel level focus accuracy only because lens is completely static when image sample is acquired. Think of it, if the lens element was moving during sample exposure the image would be blurred and it would have been impossible to tell exact position where the peak contrast was achieved. Plus things like motor vibration would blur the image like shutter shock does.

Think again. The point is not that the AF system captures a perfectly sharp image. It rarely does since focus is by definition wrong during all but a small fraction of the AF time. The point is that it determines the peak focus (max contrast) with precision as it moves along. This is merely a matter of signal strength (flatness of target, target stability within the frame, target contrast, light level) and sampling rate (how big an amount the lens AF mechanism travels between samples).

Initially AF system doesn't know position where the image will be sharp, it may occur anywhere during lens travel, so it must acquire all samples anticipating that sharpness can be close to a peak. This can't be done while lens is moving.

Of course it can.

And, at least for the last part of focusing, camera certainly can't spread exposure over multiple motor steps, otherwise the image won't be perfectly sharp.

The peak is not determined by capturing a single sharp image. The peak is determined by comparing contrast in a series of observations. If the contrast is highest somewhere and falls on either side, that's the peak.

Do a short test: Focus fast lens somewhere close on something sharp, like small text, switch it to manual focus and move focus ring just a little, couple of millimeters. The image will be out of focus. That means AF is consistently accurate with precision close to one focusing step. Cameras can only achieve that if they stop after each step and take image sample while the lens is static.

No, they can accomplish it just as well while changing focus continuously. Consider the following examples.

1. The lens moves continuously such that there is a change of 10 distance units during each exposure.

2. The lens moves 10 distance units, then stops for exposure, then moves 10 distance units again, then stops for exposure, and so on.

Why would the first example lead to lower precision than the second? In both cases, the expected margin of error is 5 units.

And how do you know whether 5 units is good enough precision? I would think that one unit is the target for accuracy. Which means lens should be static.

Whether 5 units is good enough is not the issue. Choose any number of units you please and then answer the question at issue, namely why the example where the lens moves continuously would result in lower precision than the one where it stops for exposure.

Regardless of the units, statistically, on good target, I don't think you will be able to find better focus manually than what AF finds. That means AF can't jump through many focusing steps, it has to acquire samples close enough. Initially the focusing can be rough (like during hunting) but in the end every step matters.

Again, the point is not the number of steps. See above.

The difficulty of keeping the target perfectly constant in handheld shooting, especially with longer lenses without stabilization, is in all likelihood a far stronger impediment than focus adjustments or vibration caused by the AF motor within individual AF frames.

This only means that focus may be inaccurate when handholding long telephoto lens (which is true).

This was just an example of the fact that you are not concentrating on the right thing. What matters is the variation between frames, not within frames.

Then your example is bad, which rather shows that you are concentrating on wrong things. Try again with better example.

Why is my example bad? It perfectly illustrates the point I want to make namely that the variation between frames is what matters, not the variation within frames. Do you disagree with that claim of mine? If so, on what grounds?

The problem is the comparability between frames, not within them. If the AF target suddenly becomes less contrasty from one frame to the next due to handshake rather than focus adjustment, that can obviously fool the AF system far more than motor vibrations or change of focus during a single frame.

I did a quick google search but so far only found indirect information about lens movement during AF:
"contrast detection then performs a focus fine-tune, at that distance. The 18-135mm STM's lens' design features a light, internal focus group that can be quickly accelerated and decelerated to suit this autofocus behavior"
http://www.dpreview.com/news/2012/08/15/Video-tests-of-the-Canon-EOS-650d-rebel-t4i-hybrid-autofocus-system

And you know that CDAF lenses use stepper motors for focusing, right? By definition these motors move in small steps, one step at a time. It is not uncommon for stepper motors to have a speed of thousands of pulses per second. You can't feel microsecond-long pulses. And that's why we also enjoy fast focusing speed, unless the motor doesn't have enough power to move lens focusing element fast.

Yes I know that these lenses use stepper motors and I also know how stepper motors work. The fact that the actions of stepper motors can be divided into steps, however, does not mean that they stop at each step and then start again. It merely means that you can direct them very precisely to a target which is of course an advantage. Furthermore, the important point here, as I have stressed before, is not whether the action is divided into pulses but whether, as you claimed, the AF movement is interrupted for each AF exposure, AF readout, and AF evaluation or whether these things occur in parallel, as I claim.

Try to capture an image with legacy manual lens while rotating focus ring, you will see what I mean. Set shutter speed to 1/50s, or so. When camera is focusing in low light the shutter can't be much faster.


Your theory is wrong in its claim that if the 20 mm is slower when it hunts then it must be universally slower. More specifically, your theory fails to recognize that the speed of the AF process is determined not only by the max speed of the lens AF mechanism but also by the way the body controls that mechanism. The body can allow the lens to move faster under some circumstances (e.g., when hunting) than under others and the body rather than the lens AF mechanism is often the bottleneck. Your theory does not take these facts into account.

The "theory", what you call it, is just fine. One contrast measurement per one motor movement.

No. See above.

But let's for the sake of the argument assume you are right that the AF process actually works the sequential way you think, i.e., by starting focus movement, stopping focus movement, exposing the sensor, reading the sensor, evaluating the image data, starting focus movement again and so on. Why would such a scenario necessarily imply that the contribution of the focus movement is always a constant percentage of the total AF time? And why would such a scenario necessarily rule out the possibility that the contribution is greater when the lens hunts than when it doesn't?

The contribution of the focusing element movement is a constant percentage of total focusing time when two conditions are met:
1. Exposure time is constant.
2. Camera body is not a bottleneck.

Evaluating the image data isn't part of the same sequence as lens motion. Two parallel processes are going on during autofocus:
1. Mechanical process: Lens focusing element motion, stop. Sensor exposure. Again lens motion start, motion, stop, and so on.
2. Software process: Sensor readout. Image analysis for contrast estimation. Again sensor readout and so on.

OK. So you have now revised your theory. Earlier on, you described the CDAF process as follows:

Yes, the way I described is precisely how CDAF works. Move lens, stop lens, capture image sample, measure contrast, move lens again, stop lens, capture another sample, and so on.

In your original theory the lens is stopped for exposure, sensor read-out, and data evaluation. In your revised version, it is stopped for exposure only.

I didn't revise my theory. I just explained it in details. In my original explanation I didn't say that some steps can't be done in parallel, did I? Initially I didn't have time for details such as parallelism, that's all.

BTW, it's not a theory, this is an explanation how CDAF really worked in some industrial system I worked on. It wasn't a consumer camera but I see no reasons why it would work differently in cameras that we use.

What was the focusing task faced by these systems and what was the AF speed requirements?

For some reason you chose to omit the passage of yours quoted immediately above (in yellow). For some reason you also refrained from answering the question I asked about that passage (in purple). Could you please answer that question now instead.

I omitted it for (I think) obvious reasons - our conversation grew up to ridiculous size. In fact, I was deleting irrelevant parts of it before. If I haven't done that we would have probably reached post size limit and this conversation would have collapsed. I don't think anyone except us would care, though.

If your work-related experiences with CDAF systems are irrelevant, why did you keep referring to them as some sort of evidence that you know more about such system than I do? In addition, I think it is not for you but for me to decide whether an answer to the question I raised merits an answer.

So, I hope you'll understand that I don't intend to continue this conversation.

Up to you of course.

I don't think it's as productive as it might have been (nobody is to blame, just a twist of fate, perhaps next time will be better).

Answering your question - the focusing task was to focus camera lens on some more or less contrast target. Camera had a screw motor adapted for focusing mechanism of Pentax AF lenses (which is why I was able to borrow Pentax 77mm F1.8 limited lens for photography and tests - our company still has it in the lab, left from that project).

Thanks for providing this information. I now fully understand why your work-related experiences with CDAF are indeed irrelevant but, in view of that, I have even greater difficulties understanding why you brought them up in the first place as some kind of proof of your superior expertise.

AF speed requirements were to provide reasonable fast focus speed. Our motor and this (non-CDAF) lens became a limit at some point.

I can certainly imagine that.

What I'm saying is that for understanding of focus speed limitation it's not that important whether image processing goes in parallel with lens motion

I know that you have made that claim all along, yet never backed it up.

or it takes insignificant time (close to zero). That's why I omitted parallelism in the beginning.

Are you now saying that AF exposure, read-out, and evaluation times are all insignificant? If they are, what would then explain why older MFT bodies have AF times nearly twice as long as more recent ones although the lenses used are the same?

If your best take is to point to what I omitted in some brief description - go for it. Without me, though, as I think I already explained everything.

See above.

What you initially said was exactly what I quote. What you describe in that quote is a purely sequential process. Nothing takes place in parallel according to that quote. I immediately objected to this sequential description. You didn't acknowledge that I was right. You have now revised your theory, without recognizing that you revised it.

I'm not taking an exam, I'm describing what I think matters most and what I have time for.

You could have saved some time for both of us by not doing what you did, i.e., first claim that the process is completely sequential, not acknowledging my objections to that claim, and then, much later, revise what you inititally said. But I am glad that you have opened up for at least some elements of parallelism.

I think you could have saved more of your time by not responding to every post that indicates slow focusing speed of 20mm lens. Just a suggestion.

What makes you think that I respond to posts on this forum to save time? That sounds far-fetched in the extreme.

If you want to catch me on something - you may focus on that if you think it's beneficial for you.

I am not focusing on it. I am merely pointing it out so that we have a clear view of where you stood and where you currently stand. That appears to be important to avoid misunderstanding.

Sequential versus parallel processing is at the core of the discussion between us and has been so for quite a while. Feel free to look back at prior posts. It is everything but a minor detail.

No, the core was that the lens is a bottleneck and that lens motor + focusing element mostly define focus speed, precisely the element start/stop time.

That's a claim of yours that I contested as far as normal focusing (not hunting) is concerned. I contested it largely on the ground that it is based on unreasonable assumptions about sequential processing.

Also that if focus hunting is slower then normal focusing will be slower too. This is really simple because focus hunting is just like normal focusing, only with extended focus distance range.

No it won't since the body is the bottleneck in normal focusing but the lens AF mechanism the bottleneck when the lens is hunting. Why such is the case is something I have already explained at length.

If you prefer to stick to my original explanation - consider image processing time negligible and you will see that the focusing speed will be proportional to focus motor speed.

I don't prefer to stick to your original explanation since that was even worse. In some cases (normal focusing), the lens will not travel at the maximum speed its mechanism allows since letting it to do so would yield a too low sampling rate (exposure-readout-evaluation cycles per time unit divided by the focus movement per time unit) for sufficient precision in determining the location of the peak. In this case, the lens AF mechanism will not be the bottleneck and the 20/1.7 will focus as fast as other lenses.

In other cases (hunting) it is not a problem to let the lens AF mechanism travel at its max speed in which case that mechanism will be the bottleneck.

You obviously include the time where the lens just waits for the exposure to finish, after it has stopped and before it has started, in the lens bottleneck. That's obviously incorrect. The lens could have moved during that time but the body did not allow it to do so.

The exposure time is something that can't be significantly minimized. Lens motor timings can be. As we know, some lenses are faster to focus than others in the same conditions. So, it is correct.

Are you seriously saying that the time the body tells the lens to stay where it is should be counted as part of the lens bottleneck time?

The body doesn't tell lens to stay. It tells to start and to stop. There is a difference.

I hoped that an elementary semantic analysis would suggest to you as it does to everyone else that if the body has told the lens to stop and hasn't told it to start again, it is effectively telling it to stay. Apparently, I was wrong.

Whether and to what extent something can be minimized is obviously completely irrelevant when it comes to deciding whether it should be considered a lens bottleneck or not.

The lens is waiting for the body to let it go. It's not the body waiting for the lens to reach the proper position. Hence, exposure must be moved to the other side of the equation.

I think I already said that before. It takes longer to accelerate heavier focusing element and to start/stop moving it.
Body finished image processing of previous sample and waits for lens to stop to acquire next one. Then, slower lens will accelerate slower, again delaying acquisition of the next sample. This is how the lens can be a bottleneck.
Lens isn't waiting for anything. Body waits for lens to move to the next position.

During the exposure the lens waits according to your theory. Or are you now saying that it is decelerating or accelerating during the exposure?

See above. Lens doesn't wait. Camera waits for both exposure and lens motion. Exposure time can't be reduced but the lens can be replaced with a faster one.

During the exposure, the lens certainly waits for the camera to tell it to move (based on your theory of how things work, not mine). The camera waits for itself to finish the exposure. Clearly the time required to finish the exposure cannot be considered part of the time ascribed to the lens AF mechanism, which is of course the time required to move the lens elements from one focus point to another when the body tells the lens to move, nothing else.

Exposure time could be reduced if the lens was replaced by a faster one but max aperture isn't the issue we are debating here, is it? It's the impact of the lens AF mechanism.

I meant faster focusing lens, the lens with faster motor, not the lens with faster aperture.

OK. Forget the second of the two paragraphs then. The first stands.

2. You forget that even if exposure time as well as the sampling rate (the frequency at which the body takes new exposures) is constant the body can allow the lens to move more or less (faster or slower) between exposures. One important reason why the body may not utilize the max speed of the lens between exposures is that this reduces the precision by which peak focus can be determined. If the lens is allowed to move twice as much between two consecutive exposures, this by necessity cuts precision in half, everything else equal.

I didn't forget that. I just think it is not that important.

It is important among other things because it invalidates your claims about the AF process. You are assuming that the lens moves at max speed whenever it is asked to move. I pointed out one significant reason why that might not be the case.

Consider at this point again my explanation of why the lower max speed of the AF mechanism of the 20/1.7 matters more when the lens hunts than when it doesn't:

Whenever the body has a clear signal (i.e., can clearly determine whether contrast is increasing or falling) and is on the right track (contrast is increasing), it is looking for the peak. Consequently, it cannot allow the lens mechanism to move too fast, i.e., too much between each sample, or the location of the peak will be determined with insufficient precision. When, however, the lens is hunting (lacks a clear signal) its objective is to pick up such a signal first of all. For this purpose, big focus movements between each reading is an advantage rather than a disadvantage since it makes contrast changes due to focus adjustments more obvious. Consequently, the body will allow the lens to move as fast as it can.

Yes, during hunting image samples are probably acquired a bit less often than during normal focusing.

Fine. If you agree that samples are acquired less frequently per lens distance unit travelled (not per time unit) when hunting, then you agree that the lens mechanism is not travelling at its max speed except when hunting.

I don't agree that it's important. Again, hunting is just like a normal focusing, only on extended range. Camera body may take shortcuts by reducing frequency of samples, but still, it has to acquire enough samples in order to find a point where it can stop hunting.

It is important because it vindicates my explanation of why the 20/1.7 takes longer to hunt to completion than other lenses but doesn't take longer in normal AF operations. Again, when hunting, the 20/1.7 is held back in comparison with other lenses by its slow lens AF mechanism. In normal focusing, when that mechanism isn't the bottleneck, it is about as fast as other lenses.

Still, the samples should be acquired frequently enough. The more samples, the better chances camera finds the focus point and stops hunting.


 Anders W's gear list:Anders W's gear list
Panasonic Lumix DMC-G1 Olympus OM-D E-M5 Olympus E-M1 Panasonic Lumix G Vario 14-45mm F3.5-5.6 ASPH OIS Panasonic Lumix G Vario 7-14mm F4 ASPH +21 more
Reply   Reply with quote   Complain
Post (hide subjects)Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark post MMy threads
Color scheme? Blue / Yellow