A study of the effectiveness of K7 shake reduction

[DDoram]

Great work. Pentax should hire you.

Dale
--
http://www.pbase.com/abundant108

http://www.pentaxphotogallery.com/home#section=ARTIST&subSection=272176&subSubSection=1787360&language=EN

 

[Oleg_V]

This is the best SR test I have seen so far. Though some important questions are left to be answered.

• How 'static blur' depends on the exposure.
• Why blur around 1/80 is noticeably higher than at slower shutter speeds even with SR off ? I think this is the indication of the universal character of the effect known as 'K-X mirror slap problem'

And the one big notice about the methodology in general. The difference between SR on and off positions is not the real effectiveness of the SR since in reality the SR is never turned off completely - the sensor is always floating in the magnetic field even with SR 'off'. The real effectiveness is the difference between SR on and sensor mechanically fixed which is impossible to achieve without dissembling the camera. The tests like this may give positive results even if in reality SR just adds more blur to the picture if its 'turned off'. I'm almost certain this is true within the problematic range around 1/80.

 

[labnut]

This is the best SR test I have seen so far.

Thanks. I tried to add some precision to an area of imprecise debate. Or as I like to say 'Make measurements, not assumptions'.

Though some important questions are left to be answered.

I will try to answer

• How 'static blur' depends on the exposure.

My terminology has come back to haunt me! What I called static blur is the blurring introduced by the optical properties of the system, measured while the camera is rigidly mounted, hence 'static'. This varies according to aperture in accordance with the normal optical properties of a lens. This was necessary to know because of the constant lighting intensity. For the final measurements I varied the shutter speed which produced a different aperture for each shutter speed setting. I then subtracted the static blur for the same aperture.

• Why blur around 1/80 is noticeably higher than at slower shutter speeds even with SR off ? I think this is the indication of the universal character of the effect known as 'K-X mirror slap problem'

Yes, at first that was my conclusion as well. Then I considered an alternative explanation. The blur curve without SR shows a roughly exponentially increasing blur as shutter speed decreases. With SR On the blur curve shows an identical trend until about 1/80 - 1/100 sec. From there the blur starts to fall off as SR takes effect. This would be true if SR starts taking effect at 1/100 sec and there is also a soft roll-off. So at 1/80 sec SR is only partially effective while at 1/50 sec it is fully effective. The Murata Gyrostar sensor has a 3dB bandwidth of 100 Hz so I think this is the likely explanation. But this is informed speculation. The only way to confirm this would be to conduct the handheld measurements with mirror-lockup. Perhaps I can achieve that by using the Live View feature.

And the one big notice about the methodology in general. The difference between SR on and off positions is not the real effectiveness of the SR since in reality the SR is never turned off completely - the sensor is always floating in the magnetic field even with SR 'off'.

Indeed. With SR off there will be no feedback signal instructing the sensor to move. But will the sensor remain motionless, locked in place by the magnetic field? Pentax claim this and in the absence of better information I have essentially assumed this is true.

The real effectiveness is the difference between SR on and sensor mechanically fixed which is impossible to achieve without dissembling the camera. The tests like this may give positive results even if in reality SR just adds more blur to the picture if its 'turned off'. I'm almost certain this is true within the problematic range around 1/80.

I can think of no simple way to test this directly.
But we can get a reasonable estimate of the effect. Look at Illustration No 4.
For a rigidly mounted camera at 1/80 sec:-
Vertical Image blur with MLU = 1.11 pixel (essentially the optical resolution)
Vertical Image blur, SR Off without MLU = 1.32 pixel
Vertical Image blur, SR On without MLU = 1.32 pixel

So image blur increases by 0.21 pixel at 1/80 sec without MLU. This is probably a reasonable estimate of the movement of the sensor caused by mirror slap.
Thanks for your useful comments.
Peter

 

[Oleg_V]

Thank you for the comments

The Murata Gyrostar sensor has a 3dB bandwidth of 100 Hz so I think this is the likely explanation.

IMHO its not much convincing. First, there are no source of high frequency vibration after the shutter is open. Second, the shutter speed is not related to the vibration frequency at all..

Indeed. With SR off there will be no feedback signal instructing the sensor to move.

Not at all. With SR on the SR apparatus is trying to compensate camera rotation. But with SR off its trying to stabilize sensor position with respect to the camera since its never fixed mechanically in place - its always 'levitating' in the magnetic field. You can find the details in the relevant patent http://v3.espacenet.com/publicationDetails/originalDocument?CC=US&NR=2006067660&KC=&FT=E

 

[Henry Richardson]

Thanks. Very interesting.

--
Henry Richardson
http://www.bakubo.com

 

[falconeyes]

So eventually I decided to accurately measure the effectiveness of Shake Reduction in my K7. The project just grew and grew until in the end I had completed a large study.

labnut aka P.Smith, thanks for a great study.

I have taken your results and extracted the numerical data to reinterpret the result.





in this form, it is easier to understand how blur depends on exposure time and how the SR mechanism works. I reconstructed error margins from the variance you posted and the claimed N=10 sample size.

So w/o SR, there basically is a linear dependance. With SR, there is a linear dependence too over a wide range, within a region 1/50s to 1/8s. Below 1/100s as you note, the SR has no effect and above 1/7s, it slowly seems to loose effectiveness.

Its overall effect is 3.0 stops (ratio of slopes of dotted lines) with a better efficiency (4+ stops) around 1/20s.

 

[labnut]

Thanks falconeyes. That is a better way to present the results as it shows the underlying linear relationship within the area of interest. Showing the error margins is very helpful.

On a more general note, no one commented or questioned the precision of the blur measurement to a tenth of a pixel. This is a consequence of the slant edge measurement where, by measuring along a five degree angle to the edge, one has in effect a superresolved image, allowing measurements in excess of the resolving power of the sensor.
Peter.

I have taken your results and extracted the numerical data to reinterpret the result.





in this form, it is easier to understand how blur depends on exposure time and how the SR mechanism works. I reconstructed error margins from the variance you posted and the claimed N=10 sample size.

 

[samusan]

Thanks for the efforts, really impressive quality facts and measures.

With that much time and skills available to do such a study, I'd say you must be a retired academic. Prof Smith (from Piled Higher and Deeper), is that you ?


--
Samusan

 

[Kikool]

Many thanks to both of you, real good work !

On this forum and others, we often read people debating about SR efficiency vs VR or whatever in lens stab. system, and claiming figures as generally "admitted", published by the marketing services or compiled from various "tests" or "reviews".

Every time, I want to participate and just state that such figures are rubish. At first because the testers and testing conditions are not identical. There are also so many parameters that only a real study, or comparision side by side for a couple of camera/lenses can be done, just to give one quite accurate answer in defined conditions.

That's what you did for one camera/lens. And that's just great
Thanks again for that.

Cheers
BR

 

[dlacoutu]

Labnut, I'd like to reproduce your experiment (having access to a K10, K20, Kx and K7! And possibly a *istDS, to use as a fixed sensor reference), so could we go through the exact methodology, along with how you measured your pixel blur?

I can repeat the test with the following lenses:

• DA18-55 (II and WR)
• DA 18-250
• FA50/1.4 & F50/1.7
• FA28
• Tamron 28-75 and 70-200
• Sigma 18-125 and 18-200
• various MF lenses (to check the SR accuracy with those)

Not wanting to test all of them, but that's just so we can agree on what should be tested (I've thought something like 18/50/200, then some MF lenses)

I'm in mind of testing different hand-holding techniques, also the orientation's impact on SR (I've noticed that my vertical shots are often blurred).

 

[Oleg_V]

And possibly a *istDS, to use as a fixed sensor reference

Would be extremely interesting. I'd expect istDS wins hands down at wide angle and exposure around 1/100.

 

[labnut]

Thanks for the efforts, really impressive quality facts and measures.

Thanks for your kind words.

With that much time and skills available to do such a study, I'd say you must be a retired academic. Prof Smith (from Piled Higher and Deeper), is that you ?

Not retired and no academic pretensions.

The great advantage of a name like mine is the complete anonymity it confers. I don't have to worry about libel/slander or data privacy and I am automatically immune to identity theft. There are many eminent P Smiths but I am most decidedly not one of them

 

[falconeyes]

Labnut, I'd like to reproduce your experiment (having access to a K10, K20, Kx and K7! And possibly a *istDS, to use as a fixed sensor reference), so could we go through the exact methodology, along with how you measured your pixel blur?

dlacoutu, are you aware how much work labnut must have put into this?

If my counting is correct, he measured edge blur for 840 edges!

This is why I reconstructed his numbers rather than repeated the measurement.

I don't know for QuickMTF (which isn't free but affordable). But with SFRedge (which IS free) it would take quite a moment to measure all edges.

Please keep in mind that an automated procedure would have to find the proper ROIs in the hand-held images. Feasible but most testing suites don't do it!

As for precision of edge blur measurements: I found a systematic error in the order of 0.15px which is quite good. Results are useful down to 0.05px but with larger errors than statistics would imply.

 

[Roland Karlsson]

Thanx!!

I have been wondering for some time if anyone had done a test like this.

Its all good and cosy to know that your 1/4 s images goes from total blur to just some blur - but - whats happening for short exposure times? Are there any risks for worse images at 1/500?

Your investigation shows this not to be the case. Nice!

And do you get improvements at 1/50 and 1/100?

1/50 - yes. 1/100 - not with this lens really.

I am somewhat surprised by one thing though. You are getting quite "reliable" data at high speeds. There you show resolutions at 1/10 of a pixel. But - the lens do not resolve that much. The lens resolution is less than one pixel. It sounds to me that computing the difference between dynamic and static blur would be less predictable and sometimes even negative.

Now - personally I would have avoided all such problems by using a longer focal length. Then the blur (without OS) would be larger. POK - its not the same test. The OS algorithm is probably modified when changing focal length.

--
Roland

support http://www.openraw.org/
(Sleeping - so the need to support it is even higher)

X3F tools : http://www.proxel.se/x3f.html

 

[dlacoutu]

We could also add the laser stylus trick, in order to really measure the efficiency of different handholding techniques, and everybody could see what my hands are (in)capable of.

Falconeyes, I'm aware of the commitment involved...

I've spent whole nights testing my various focus screens at every aperture, for every lens I had, to map their response and understand the K10/K20 exposure problems.

I've experienced some weird blurs with my K7 on some occasions, and I'd really like to understand why and when.

For instance, on one occasion, I had repeatably blurry shots (camera shake, not OOF) at 1/180 with a 28-75 set to 63mm, vertical orientation. Couldn't take one sharp shot, on at least 10 tries!!! I just gave up, perplexed.
Another occasion, 18-55 set to 38mm, 1/100, vertical. Got one sharp in 5 shots.

I had no such problems with my K10 & K20 with the same lenses, so I'm rather inclined to think I'm not the culprit here...

 

[falconeyes]

So eventually I decided to accurately measure the effectiveness of Shake Reduction in my K7.

I already commented about my re-interpretation of labnuts work.

This time, I tried to understand the shake-induced blur in a more mathematical way.

To this end, I used a LOG-LOG plot:





Let b be the blur width (in µm), a and t0 be constants, f be the focal length (in mm) and t be the exposure time (in ms).

Then, P.Smith's data are, within error margins, fully compatible with this formula (dotted lines in the above plot):

b = a f |t - t0|

With values as follows:

SR OFF:
t0 = 0
a = 1 / (280 s)

SR ON:
t0 = 44 ms
a = 1 / (1200 s)
b = MIN (b_sroff, b)

This corresponds to an advantage of 2.1 stops far from t0 and more than 4 stops around t0.

The crossover point is at about 8 ms (1/125s) where SR becomes ineffective.

It will be most interesting to study if this formula changes with the focal length. I imagine it doesn't which would mean that the SR in the K-7 is less useful with focal lengths in excess of 150mm and that even longer focal lengths should be used at 1/25s or using the 1/f rule.

 

[Roland Karlsson]

I already commented about my re-interpretation of labnuts work.

This time, I tried to understand the shake-induced blur in a more mathematical way.

To this end, I used a LOG-LOG plot:
...

Hmmmm ... the static blur is in excess of one pixel. I would be very careful to put any relevance to measurements thats smaller than one pixel, i.e. most of your graph.

--
Roland

support http://www.openraw.org/
(Sleeping - so the need to support it is even higher)

X3F tools : http://www.proxel.se/x3f.html

 

[falconeyes]

falconeyes wrote:

Hmmmm ... the static blur is in excess of one pixel. I would be very careful to put any relevance to measurements thats smaller than one pixel, i.e. most of your graph.

I am aware of this. Yet, I beg to differ.

P.Smith's method is capable to measure subpixel blur and he subtracted the static blur. Static blur is due to the finite pixel size, demosaicing and lens blur. Whatever it is, one has a second measure in the linear dependency of blur vs. shutter which has static blur as its offset.

If you look at my graph, you see that systematic errors start to be significant at about 0.15px and results start to be random below 0.05px. Which is good enough to draw my conclusion.

 

[falconeyes]

b = a f |t - t0|

SR ON:
t0 = 44 ms
a = 1 / (1200 s)

It will be most interesting to study if this formula changes with the focal length. I imagine it doesn't which would mean that the SR in the K-7 is less useful with focal lengths in excess of 150mm and that even longer focal lengths should be used at 1/25s or using the 1/f rule.

I've run informal tests with a 300mm lens, at 1/320s, 1/160s, 1/100s, 1/25s, SR ON and OFF.

The blur didn't seem to depend on SR ON or OFF for 1/320s, 1/160s, 1/100s and blur was less at 1/25s with SR ON but still a little bit more than at 1/160s. This is all very preliminary and not academic!!!

It is just an indication that the formula above may indeed not depend on the focal length.

Which would mean (if the first quick test was valid at all) that the Pentax SR is very good for low light photography but not so good for long range tele photography.

Therefore, a telelens-based optical stabilization (like in the new Sigma 50-500) may make a lot of sens even with a current Pentax body.

More tests are required to verify or falsify the above statement, though!!

 

[Roland Mabo]

Or is the SR system in the K-7 pretty much the same as those other Pentaxes?

K-7 has improved effeciency with faster response time - thank's to stronger magnets and improved SR analysis algorithms. This in comparision with the K20D. The SR system in the K-7 is more effective (according to Pentax) than in other Pentax DSLR's.
--
Take care
R
http://www.flickr.com/photos/raphaelmabo