One of the most acclaimed new features on the E-510 is the new system of image stabilization. Unfortunately, there is not much known about how effective it may be. The data available on the Web range from anecdotal evidence at best to wishful thinking and numbers taken off the ceiling at worst.

The situation is not helped by the various available software, intended to "measure" the advantages of the system. No software will replace a good experiment design and a proper analysis of results, and you cannot use the software without knowing what you are doing: some of the results I've seen remind me very much of a monkey playing with a computer: yes, it knows how to press the keys. Like line plots, graphing a measure of camera shake versus frame number — it's beyond entertaining, just sad.

It is easy to criticize, less so — to do things right, or, at least, almost right. This is hampered at the very start by the fact, that the most frequently asked question: "How many F-stops of advantage do we get from the IS?" is not the best statement of the problem.

The effect of camera shake is not a binary one: out of a number of handheld pictures taken under the same conditions, each may have a different amount of shake effect visible. The movement of our hands, holding the camera, is a random effect, and the effectiveness of the camera's detection and correction system may be different in each case. Therefore the experiment has to be designed accounting for the statistical nature of the process, at the same time avoiding (as much as possible) introduction of any systematic factors.

The photographer's shooting technique is one of the major components here. Some of us will be able to handhold longer shutter speeds without IS than some others with it. Your body position and breath control, for example, may affect this greatly. Then the camera's mass (inertia!), the same about the lens weight and balance. The lens' focal length: usually we assume that the magnitude of the visible shake effect is directly proportional to it, but this is true only if for various lenses the angular instability of the camera remains the same (which is not the case).

Before you've invested your precious time into reading this article, this is the last warning: if you are looking for a simple answer: "this IS system gives you 2 stops of advantage, compared to camera's A 1.5 stops and Camera's B 3 stops", stop reading and look elsewhere.

You may also find articles which explain the general theory or relativity in three paragraphs and two colorful pictures. You will feel better, without really knowing more.

If you decide to stay, be prepared to see some minutiae. I'm going to describe the whole process in detail, so that no part is hidden under a "trust me" cover.

Experiment design

A high-contrast target with lots of detail (my computer LCD screen) was photographed handheld from a distance of about one meter at shutter speeds ranging from 1/2 s to 1/125 s in 1 EV increments (shutter priority AE). This was done at three focal lengths: 14 mm, 42 mm, and 150 mm, using the E-510 kit lenses; for the first two values the 14-42 mm ZD ED was used, and for the third one — the "new" 40-150 mm ZD ED. For each shutter speed and focal length combination 20 exposures were made with image stabilization and another 20 without.

The value of 20 was arrived to based on the results of my experiment of two weeks earlier, limited to one shutter speed. In that experiment I was able to check the magnitude of subjective image fluctuations, arriving to the conclusion that less than 20 would not provide enough of statistical accuracy; more would be desirable, but requires more work (and I do not want to kill my shutter just testing the camera).

I ended up with 31 data points (20 frames each), the total of 620 frames. This is what the whole analysis is based upon, and I dare say this is barely adequate; I would hold in suspicion any conclusions drawn on a smaller volume.

To preempt questions: not all focal length and shutter speed combinations within the limits mentioned above had to be checked: for example, if I reached 100% of "good" frames at 1/60 s (IS on) at a given focal length, going through a sequence at 1/125 s would be just wasteful.

All shots were done sitting, with my elbows not supported; the chair stayed in the same position. I was trying to control my breath and keep all subjective factors as close as possible. The order in which the 20-frame series were shot was random, following a previously prepared schedule, with regular breaks to alleviate any results of fatigue.

Sample evaluation

The image files were transferred to a computer and renamed so that a name identified the condition under which a given picture was taken. After that, my image browser sorted them in byte size order, with no file information displayed, so I wasn't able to say which image is which (except, of course, the focal length, easy to tell). This was intended to remove any subjective bias in image evaluation.

Every image was then submitted to visual inspection in full pixel size and categorized as "good", "medium" or "bad", depending on the visible degree of camera shake. Each then was moved to one of subfolders corresponding to these categories.

The categories were assigned subjectively, but, due to the anonymous character of the process, the same way for every shutter speed, IS on or off. Any statistical fluctuations in the evaluation will, of course, affect the statistical (again) accuracy of my results but will not introduce a systematic bias, and that's what counts.

Instead of verbally describing my criteria for "good", "medium", and "bad", let me show you examples of each (F=150 mm):

The LCD structure in the photographed text is quite helpful in detecting a slight shake, in this case vertical, of about 1/2-pixel effective magnitude (screen pixel, not camera's!). The "bad" sample is, by the way, quite close to "medium", I've seen much worse.

Numerical results and data reduction

As a result of the above, I had three numbers, one corresponding to each of these three quality categories, for every combination of focal length, shutter speed, and IS on/off.

Boring you with these values now makes no sense, but Appendix One contains the original values for those who may be more fluent than I am in data reduction and statistics, in case they would prefer to do their own number-crunching. The table also contains the values of my G metrics, described below.

To make further analysis easier, I translated each triplet to a single subjective goodness score, G, expressed as the percentage of "good" frames in the total, with two "medium" additionally counting as one "good". Thus, a series of all 20 "good" frames has a goodness score of 100% (20/20), one consisting of 10 "good" and 6 "medium" a 65% ((10+6/2)/20), and one all "bad" — a zero.

In other words, we can say that all shutter speeds and focal length combinations with G=100% are perfectly handholdable (under my conditions, at least), while those with G=0 do not give you a reasonable chance of getting a stable image. The most interesting things, however, are happening in-between, and that's where the next stage of data reduction kicks in.

Before we go deeper into that, notice that the green points are shifted systematically to the left of the yellow ones. This indicates that image stabilization, indeed, is not our wishful thinking but an experimentally confirmed fact. The next section will prepare the ground for answering the key question: how well does it work.

The goodness score model

Obviously, at shutter speeds slow enough the goodness will be at 0; at speeds fast enough — 100%; you do not need me to tell you that. The behavior in-between these two extremes can be approximated as a linear function (goodness against EV, or -log₂t, with t being exposure time in seconds).

Why linear? Because not knowing the exact shape we should use the simplest model available. In a log scale, because our data points are equidistant in this representation. The simpler the better.

Instead of drawing a line by hand, subject to human error and wishful thinking, we can use a numerical procedure of function fitting which will, out of all possible such S-shaped functions, find the one best fitting our data points. (Feel free to skip the fine print below if you are not into numerical analysis).

Our goodness-versus-EV function has a simple shape: 0 for EV less than some value, v₀, 100% for EV above another value, v₁₀₀, and a straight line connecting (v₀,0) with (v₁₀₀,100) in-between. The problem is in finding the values v₀ and v₁₀₀, for which the line will be running "closest" to the data points.

I decided to use "closest" in terms of least squares here: the "best" v₀ and v₁₀₀ will be such that the sum of squared deviations of the data points from the function value, measured in the vertical direction, will be at minimum. This is a common procedure in sciences, although it is theoretically justified only when the errors in y are distributed normally with the same standard deviation. This is not the case here, so that the least-squares fit will be a bit bastardized, but this transgression is a common thing in sciences; let any physicist not guilty of this sin cast the first stone and I'll accept it gracefully.

The above holds for the focal length of 42 mm, a given camera/lens weight, and my average shooting technique. Anyone who claims to give you an answer without these caveats is just lying (or ignorant).

Note that my 100% handholdable shutter of 1/40 s is exactly twice as long as the one often assumed, in the Four Thirds system being equal to one over twice focal length (for a 35 mm camera: one over focal length). Also note that I have a 50% chance of getting away with 1/15 s at this focal length. I used to be better 20 years ago...

Hallelujah! All we need now is to draw a model function for each focal length with and without image stabilization, and the horizontal spacing between the IS on/off lines is what we need: the IS gain, expressed in terms of EV.

Note, however, that this spacing may de different for various goodness values: the lines may not be parallel. In other terms, image stabilization may have a different effect on v₀ than on v₁₀₀, and this may be not just because of the statistical nature of the process; there may be some real reasons behind it.

To avoid hairsplitting and to further simplify the interpretation of results, let us introduce another parameter: v₅₀ — the EV value at which the goodness score is 50%. It so happens (a linear model!) that v₅₀ can be also computed by averaging v₀ and v₁₀₀. And now, the difference between the v₅₀ values with and without image stabilization is as close as we can get to answering the original question: what is the IS gain in terms of handholdable shutter speeds?

If you came here from the E-3 IS article, this is a time to go back.

The results

Let me show you the results without going through all numbers (these can be found in Appendix Two, if you care). Again, green points and lines are with image stabilization, yellow ones — without. You may visualize the EV gain as the distance between the points where an imaginary 50% line crosses the green and yellow ones.

Re-running the measurements for the outlying data point would be plain wrong and dishonest; why not do it many times, stopping only when the stochastic process brings a result satisfying our theories? For my American Readers this may be remindful of the election recount in Florida, or any other similar recount, regardless who calls it.

Anyway, that's why I wanted to use some aggregate metrics, based on more than just a few readings. With the stochastic nature of image stabilization, and with subjective evaluation, this is the only chance of getting a meaningful result.

One more remark: the image stabilization seems to be doing its best job in preventing a strong camera shake effect at longest focal lengths: not only do the 1/2 s exposures at 42 mm and 150 mm get similar goodness scores, but the v₀ values are almost identical. The inverse proportionality rule to the focal length would let me expect a 1.8 EV difference: log₂(150/42); this is well beyond any statistical error.

Conclusions

First of all, there is no doubt that image stabilization works in the E-510.

Secondly, its effects seems to be increasing with the focal length used. It remains unclear if and how it may depend on physical characteristics (mass, size) of the lens. If that dependency is significant, my numerical results will be applicable only to the lenses I've used in the experiment.

Third, as the process is based on frequency analysis of the detected camera shake (Olympus says from seven down to below one hertz), its effectiveness will depend on the nature of that shake: various users may see various gains. I think the more savvy ones, knowing how to hold the camera and trip the shutter, may gain less. I'd like to believe I'm in this category, so I wouldn't be surprised to see Aunt Minnie gaining more. Still, I have no evidence to support that theory.

Fourth and last: do not understand my reported 1-2 EV benefit as "only" 1-2 EV. Some camera makers may claim 3-4 EV (so does Olympus), but until this verified using the same measurement method and the same data reduction/interpretation scheme, it is an old wives' talk. The same as "unsurpassed color accuracy" or "mind-boggling detail".

If someone offers me his/her camera of another brand to put through a 600 frames of torture, I may consider, just for the sake of curiosity, spending a better part of a day to run a similar experiment, even if I have a (long) list of better things to do.

This is the end of my little study. What follows is two Appendices, containing the raw data and computed values of my v coefficients — of interest only to a very few Readers. Most probably you will want to ignore them, and justly so.

Appendix One: Raw data

This table show the "bad-medium-good" frame counts for each series of frames, shot at a given focal length and shutter speed combination. Below each triplet of counts, there is my goodness score in the "raw" 0..40 range (values shown in graphs above were re-normalized to percentages, i.e., multiplied by 2.5).

Appendix Two: The v coefficients

These are the values of v₀ and v₁₀₀ in terms of EV as defined in the Goodness score model section above, obtained by the least-squares fit of data point with a three-segment linear function. Values of v₅₀, also shown, have been computed simply by averaging of the two former ones. I'm showing more decimal digits than really needed, so that these values can be used in further calculations.

For a better readability, the corresponding shutter speeds (arbitrarily rounded) are shown below.

The last column shows the change in v₅₀ due to image stabilization, and the corresponding exposure time increase.