Theoretical limits of IBIS away from the center of an image

Started Sep 19, 2020 | Discussions
Kenneth Almquist Regular Member • Posts: 162
Theoretical limits of IBIS away from the center of an image
7

IBIS attempts to correct for camera motion by moving the sensor. This cannot do a perfect job of correcting for camera motion. Indeed, the whole reason that some photographers bother with shift-tilt lenses is that shifting the sensor doesn't have the same effect as changing the direction that the lens is pointing.

In this post, I will use a simple model of what IBIS does. To keep things simple, I will assume that whatever is being photographed is extremely distant, for example stars in the night sky. That means that IBIS has to compensate for changes in the orientation of the camera but not changes in its position. It also means we don't have to worry about changed to the focal plane, since everything is so far away that it will be in focus if the camera is focussed at infinity.

Here is our camera with a lens and image sensor. The colored lines are two rays from the subject being photographed.

During the course of the exposure, the direction that the camera is pointed changes:

The rays now hit different points on the sensor than they did in the first image.

The camera is rotating, but it will be easier to understand what is happening if we pretend that the camera is standing still and the universe is rotating around it. This allows us to combine the two diagrams above into a single diagram:

If you look closely, you may note that tilting the camera has a larger effect near the edge of the sensor that it does at the sensor, both because the distance between the lens and where the ray strikes the sensor is larger, and because the ray is hitting the sensor at an angle.

To put some numbers on the above effect, we need to choose a few parameters. We make the point where the blue ray hits the censor 18mm away from the center of the sensor. (On full frame cameras, the corners are about 21.6mm away from the center, so this distance will cover most of a full frame sensor). We set the level of acceptable motion blur to 0.012mm, which is two pixels on a 24 MP full frame camera.

Now I will present the equations I used to calculate the positions where the rays hit the sensor. If you don't like math, you can skip down to the table below where I provide numerical results.

In the following equations,
f is the focal length of the lens,
x is the amount that the camera is rotated,
r0 and b0 are where the red and blue rays strike the sensor before the rotation,
r1 and b1 are where the red and blue rays strike the sensor after the rotation.

r0 = 0
r1 = f * tan(x)
b0 = 18
b1 = f * tan(atan(b0/f) + x)

The initial angle of the blue ray is atan(b0/f), and we add x to that angle to compute the position after the rotation is performed.

If IBIS is enabled, then we assume that the IBIS does a perfect job of shifting the sensor so that the center of the image does not move. r2 and b2 are where the rays strike the sensor after rotation when IBIS is enabled.

r2 = r0 = 0
b2 = b1 - r1

What we are actually interested in is not the positions, but how big a rotation we can perform without exceeding our limit of 0.012mm on motion blur. To determine that value at the center of the sensor, we set the difference in position before and after the rotation (i.e. r1 - r0, b1 - b0, or b2 - b0) to 0.012 and solve for x. This gives us the following values, which I've placed in an image because the comment system doesn't seem to do tables:

The second and third columns are calculations with IBIS disabled. The second column shows the maximum angle the camera can be tilted during the shot if we want to limit motion blur to 0.012mm. The third is the maximum amount of tilt if we want to limit blur to 0.012mm everywhere within 18mm of the center. The column is labeled “at 18mm,” but 18mm is the worst case, so anything within 18mm from the center will have no more than 0.012mm of motion blur.

In our model, IBIS cancels out all motion blur at the center of the image, so with IBIS enabled we only have a column for 18mm away from the center. This improves with longer focal length.

The next column (column 5) is the one that inspired this exercise. It shows the ratio of the two preceding columns, indicating how much improvement IBIS can provide. I had expected the ratios for 20mm and 50mm to be larger.

The table shows that IBIS can be quite effective in principle with a 200mm lens, but the IBIS system would have to work fairly hard to achieve that. The final column shows how far the IBIS system has to move the sensor to compensate for the change in angle shown in column 4.

At this point, I should mention two ways an IBIS system could do better than the above table shows. First, I assume that the IBIS system attempts to eliminate all motion blur at the center of the sensor. As far as I have been able to determine, this is what all current IBIS implementations do. In principle, an IBIS system could compromise between the center and the edges, allowing some motion blur at the center in exchange for less motion blur near the edges.

Second, I have assumed that the lenses have no geometric distortion. A certain amount of barrel distortion should improve matters, while pincushion distortion will make things worse. I haven't investigated the magnitude of this effect.

In the table, I list improvement in terms of angle, but IBIS performance is normally specified in number of stops improvement in shutter speed. It's unclear how these two measures relate, but I'd question whether IBIS can achieve shutter speed improvements that are vastly better that the angle improvements that I calculate in this post.

As an anecdotal data point, let's take the Nikon 14-30mm f/4 on a Z7. The model shows an improvement of 1.604x for the 14mm focal length and 3.769x for the 30mm focal length. Ken Rockwell's review of this lens reports 1/3 of a stop improvement at 14mm and 2 stops improvement at 30mm. In this case, there is less than 1/3 of a stop difference between the angle improvements shown by the model and the actual observed improvement.

So what about the CIPA numbers showing 5 stops improvement on the Fuji X-H1 with a 14mm f/2.8 (full frame equivalent 21mm), or 7 stops improvement on the Canon R5 with the 50mm f/1.2? The first thing we note about the CIPA test is that their threshold for unacceptable blur is 0.063mm on a full frame camera, or 10.5 pixels on a 24 Mpixel camera. This less stringent standard actually makes it harder for IBIS to improve things, although the difference is minor except at 200mm:

The CIPA measurement process uses software developed by CIPA. Unless I missed it, CIPA doesn't say which portions of the image it examines, but the test chart provides some clues. The chart includes copies of a “natural image” (a picture of some fruits and vegetables in a basket), but I assume that the software ignores these, based on a comment that these are to make focusing easier. Ignore these, and the chart is a black and white checkerboard pattern, but only four boxes are guaranteed to be in the field of view (depending on the aspect ratio of the camera). So it would appear that the software is analyzing the blur of a horizontal and a vertical line, both of which pass through the center of the image. This means that the software is going get a result that is a lot closer to the blur at the center of the image than to the worst case blur anywhere on the image.

In other words, the CIPA stabilization scores are more a measure of how well the center of the image is stabilizes than a measure of how well the entire image is stabilized.

Conclusion:

1) Unless you only care about motion blur near the center of your images, the ability of IBIS to compensate for motion blur is less than you might expect on both wide angle and normal angle (50mm full frame equivalent lenses).

2) Again, unless you only care about motion blur near the center of your images, you shouldn't use the CIPA numbers when deciding which stabilized wide angle lens to buy. Lens based stabilization (unlike IBIS) may do a good job of eliminating motion blur across the frame rather than just in the center, but the CIPA numbers won't tell you this.

 Kenneth Almquist's gear list:Kenneth Almquist's gear list
Nikon D7200 Fujifilm X-H1 Nikon AF-S DX Nikkor 35mm F1.8G Nikon AF-S DX Nikkor 18-300mm F3.5-6.3G ED VR Fujifilm 16-55mm F2.8R LM WR +2 more
Canon EOS R5 Fujifilm X-H1 Nikon Z7
If you believe there are incorrect tags, please send us this post using our feedback form.
(unknown member) Veteran Member • Posts: 5,193
Re: Theoretical limits of IBIS away from the center of an image

Kenneth Almquist wrote:

IBIS attempts to correct for camera motion by moving the sensor. This cannot do a perfect job of correcting for camera motion. Indeed, the whole reason that some photographers bother with shift-tilt lenses is that shifting the sensor doesn't have the same effect as changing the direction that the lens is pointing.

In this post, I will use a simple model of what IBIS does. To keep things simple, I will assume that whatever is being photographed is extremely distant, for example stars in the night sky. That means that IBIS has to compensate for changes in the orientation of the camera but not changes in its position. It also means we don't have to worry about changed to the focal plane, since everything is so far away that it will be in focus if the camera is focussed at infinity.

Here is our camera with a lens and image sensor. The colored lines are two rays from the subject being photographed.

During the course of the exposure, the direction that the camera is pointed changes:

The rays now hit different points on the sensor than they did in the first image.

The camera is rotating, but it will be easier to understand what is happening if we pretend that the camera is standing still and the universe is rotating around it. This allows us to combine the two diagrams above into a single diagram:

If you look closely, you may note that tilting the camera has a larger effect near the edge of the sensor that it does at the sensor, both because the distance between the lens and where the ray strikes the sensor is larger, and because the ray is hitting the sensor at an angle.

To put some numbers on the above effect, we need to choose a few parameters. We make the point where the blue ray hits the censor 18mm away from the center of the sensor. (On full frame cameras, the corners are about 21.6mm away from the center, so this distance will cover most of a full frame sensor). We set the level of acceptable motion blur to 0.012mm, which is two pixels on a 24 MP full frame camera.

Now I will present the equations I used to calculate the positions where the rays hit the sensor. If you don't like math, you can skip down to the table below where I provide numerical results.

In the following equations,
f is the focal length of the lens,
x is the amount that the camera is rotated,
r0 and b0 are where the red and blue rays strike the sensor before the rotation,
r1 and b1 are where the red and blue rays strike the sensor after the rotation.

r0 = 0
r1 = f * tan(x)
b0 = 18
b1 = f * tan(atan(b0/f) + x)

The initial angle of the blue ray is atan(b0/f), and we add x to that angle to compute the position after the rotation is performed.

If IBIS is enabled, then we assume that the IBIS does a perfect job of shifting the sensor so that the center of the image does not move. r2 and b2 are where the rays strike the sensor after rotation when IBIS is enabled.

r2 = r0 = 0
b2 = b1 - r1

What we are actually interested in is not the positions, but how big a rotation we can perform without exceeding our limit of 0.012mm on motion blur. To determine that value at the center of the sensor, we set the difference in position before and after the rotation (i.e. r1 - r0, b1 - b0, or b2 - b0) to 0.012 and solve for x. This gives us the following values, which I've placed in an image because the comment system doesn't seem to do tables:

The second and third columns are calculations with IBIS disabled. The second column shows the maximum angle the camera can be tilted during the shot if we want to limit motion blur to 0.012mm. The third is the maximum amount of tilt if we want to limit blur to 0.012mm everywhere within 18mm of the center. The column is labeled “at 18mm,” but 18mm is the worst case, so anything within 18mm from the center will have no more than 0.012mm of motion blur.

In our model, IBIS cancels out all motion blur at the center of the image, so with IBIS enabled we only have a column for 18mm away from the center. This improves with longer focal length.

The next column (column 5) is the one that inspired this exercise. It shows the ratio of the two preceding columns, indicating how much improvement IBIS can provide. I had expected the ratios for 20mm and 50mm to be larger.

The table shows that IBIS can be quite effective in principle with a 200mm lens, but the IBIS system would have to work fairly hard to achieve that. The final column shows how far the IBIS system has to move the sensor to compensate for the change in angle shown in column 4.

At this point, I should mention two ways an IBIS system could do better than the above table shows. First, I assume that the IBIS system attempts to eliminate all motion blur at the center of the sensor. As far as I have been able to determine, this is what all current IBIS implementations do. In principle, an IBIS system could compromise between the center and the edges, allowing some motion blur at the center in exchange for less motion blur near the edges.

Second, I have assumed that the lenses have no geometric distortion. A certain amount of barrel distortion should improve matters, while pincushion distortion will make things worse. I haven't investigated the magnitude of this effect.

In the table, I list improvement in terms of angle, but IBIS performance is normally specified in number of stops improvement in shutter speed. It's unclear how these two measures relate, but I'd question whether IBIS can achieve shutter speed improvements that are vastly better that the angle improvements that I calculate in this post.

As an anecdotal data point, let's take the Nikon 14-30mm f/4 on a Z7. The model shows an improvement of 1.604x for the 14mm focal length and 3.769x for the 30mm focal length. Ken Rockwell's review of this lens reports 1/3 of a stop improvement at 14mm and 2 stops improvement at 30mm. In this case, there is less than 1/3 of a stop difference between the angle improvements shown by the model and the actual observed improvement.

So what about the CIPA numbers showing 5 stops improvement on the Fuji X-H1 with a 14mm f/2.8 (full frame equivalent 21mm), or 7 stops improvement on the Canon R5 with the 50mm f/1.2? The first thing we note about the CIPA test is that their threshold for unacceptable blur is 0.063mm on a full frame camera, or 10.5 pixels on a 24 Mpixel camera. This less stringent standard actually makes it harder for IBIS to improve things, although the difference is minor except at 200mm:

The CIPA measurement process uses software developed by CIPA. Unless I missed it, CIPA doesn't say which portions of the image it examines, but the test chart provides some clues. The chart includes copies of a “natural image” (a picture of some fruits and vegetables in a basket), but I assume that the software ignores these, based on a comment that these are to make focusing easier. Ignore these, and the chart is a black and white checkerboard pattern, but only four boxes are guaranteed to be in the field of view (depending on the aspect ratio of the camera). So it would appear that the software is analyzing the blur of a horizontal and a vertical line, both of which pass through the center of the image. This means that the software is going get a result that is a lot closer to the blur at the center of the image than to the worst case blur anywhere on the image.

In other words, the CIPA stabilization scores are more a measure of how well the center of the image is stabilizes than a measure of how well the entire image is stabilized.

Conclusion:

1) Unless you only care about motion blur near the center of your images, the ability of IBIS to compensate for motion blur is less than you might expect on both wide angle and normal angle (50mm full frame equivalent lenses).

2) Again, unless you only care about motion blur near the center of your images, you shouldn't use the CIPA numbers when deciding which stabilized wide angle lens to buy. Lens based stabilization (unlike IBIS) may do a good job of eliminating motion blur across the frame rather than just in the center, but the CIPA numbers won't tell you this.

Interesting !!!

I had a doubt at the beginning but I think you are perfectly correct. The rotation can not be compensated exactly all across the image by moving the sensor...

Added to that, a translation move can be compensated only for the focus plane...

So neither rotation nor translation are perfect.

I am also interested in IBIS and I do regret many aspects of IBIS are not explained. Looks like nobody cares.

Another important point is to know when IBIS is really in action. Does it start when you press rhe shutter release, which means it is centered ? Or is it always in action (because you may need it during composition) ? This is a VERY important point.

This is why I disagree a bit when you say it is more difficult to compensate for 200m FL. The shutter speed is proportionnally faster when you increase FL so the move needed to compensate is the same for whatever FL.

Some camera makers (panasonic for instance) did pretend it was not working for long FL, this is a misconception. This has been proven wrong by many IBIS systems which show very good performance at ong FL

It depends a bit if you want to benefit from IBIS during composition also, in this case it depends on FL... I guess this is the reason they use both lens stabilisation and sensor stabilisation. If I had to design the system I woud use lens stabiisation during composition to have the sensor centered.

Sorry to extend the topic of your thread, my point is that I really regret there are no more details about how IBIS works, it coud be really interesting. Olympus has several modes for IBIS but for the other camera makers, this is a mysterious.

kenw
kenw Veteran Member • Posts: 6,932
Tremors are oscillations - not monitonically increasing...
4

So first this is a nice analysis, thanks so much for sharing it!

I think you are really onto something here but I also think your conclusions may have significantly overstated the case. They are implicitly based on an assumption that IS performance is limited by the maximum absolute angular offset that can be compensated for. This is equivalent to a model in which the hand motion being compensated is a monotonically increasing angular error. That's not how hand tremor works though.

Hand tremor is an oscillation with a spectral power that increases with decreasing frequency. Here are some examples:

From https://www.st.com/content/ccc/resource/technical/document/white_paper/c9/a6/fd/e4/e6/4e/48/60/ois_white_paper.pdf/files/ois_white_paper.pdf/jcr:content/translations/en.ois_white_paper.pdf

Note here that for instance that using 1/f as the reference shutter time for a FF hand held exposure that taking 20mm and thus 1/20 second down by five stops gets us to about 1.5 seconds or 0.66 Hz.  And note that the mean magnitudes at that frequency are about 0.05 degrees.  You determined a limit at the 18mm offset of about 0.04 degrees so this seems to imply that in fact five stops IBIS compensation is achievable even at an extreme wide angle.

So I think your column 5 ratio isn't really directly related to shutter speed improvement limits at all in the real world and the system can greatly exceed what is implied by that.

As to the second conclusion that OIS would be better this overlooks an issue with OIS - it can't compensate for rotation around the optical axis.  This can be a significant contribution to total blur from the physical shutter press of the finger causing the camera to rotate around the lens optical axis (or roll axis if you think of the camera as an airframe).  This is one reason without IBIS it is recommended to use a delayed shutter to improve hand held shooting.

In general the actual limitation in IBIS for normal to wide focal lengths is I believe gyro drift for measuring the motion.  For longer focal lengths IBIS motion rates and IBIS motion limits come into play I think, one reason for a preference for OIS at long focal lengths.  It looks like you may be identifying an additional constraint on the ultra-wide angle side of things which I haven't seen analyzed before.

Thanks a bunch for sharing this!  Again while I think your conclusion may be overstated it does sure seem that as we approach ultra-wide-angle focal lengths we are getting into the range where the typical absolute magnitude in hand tremor could cause a difference in effectiveness between the center and the edges/corners.

FWIW I was inspired to do some quick tests with a Z7 and 14-30/4 at 14mm but was not able to come up with any conclusive examples of this happening at the boundary of IBIS performance (i.e. I couldn't find an example with an obivously blurrier corner than the center).  This was a quick test of only a few shots at each shutter speed with little care in the setup so I don't think there is anything to be drawn from it really.

-- hide signature --

Ken W
See profile for equipment list

 kenw's gear list:kenw's gear list
Panasonic Lumix DMC-GM1 Nikon Z7 Panasonic Leica DG Summilux 15mm F1.7 ASPH Nikon Z 14-30mm F4 Nikon Z 24-200mm F4-6.3 VR +42 more
(unknown member) Veteran Member • Posts: 5,193
Errors not neglectible at all... For wide angle

Hello,

The equations are not too complicated and I have the same conclusion as the op, the error can be significant

Let:

  • x be the distance from the center of point P on sensor
  • a the angle of rotation
  • t the movement from the center corresponding to the rotation ( we have tan(a) = t/ FL
  • fl focal length

The point P, after rotation has this new distance from the center:

Then the new position on the sensor is: (t + x) / (1 - x * t/fl^2)

We can consider than the sensor compensates with a move = t (with compensates precisely the center)

Then the error is: error(x, t) = (x+t) * (1 / (x * t/fl^2) - 1)

if we take fl = 14mm, x = 17.5mm (for FF, in the extremity) If we move only 1 pixel (t=0.006mm) Then the error is 0.009mm, more than 1 pixel !

Without IBIS it would move 0.015mm = 2.5 pixels instead of 1 pixel in the center. IBIS reduces it to 1.5 pixels but this remains significant.

Demo for the equation: After rotation, point P at position (fl, x) is at this new locaiton:

  • fl cos(a) - x sin(a)
  • fl sin(a) + x cos(a)

We have to project it on plane at distance FL Then the position from center is:

(fl sin(a) + x cos(a)) fl / (fl cos(a) - x sin(a))

This is equal to: (fl tan(a) + x) fl / (fl - x tan(a))

Or (t + x) / (1 - x * t/fl^2)

(unknown member) Veteran Member • Posts: 5,193
Re: Tremors are oscillations - not monitonically increasing...

kenw wrote:

As to the second conclusion that OIS would be better this overlooks an issue with OIS

it can't compensate for rotation around the optical axis.

For translation moves, but neither can IBIS in fact.

If the position of the camera moves then this is a different perspective and nothing can really compensate for it.

IBIS compensates just for one plane at the focus distance.

FWIW I was inspired to do some quick tests with a Z7 and 14-30/4 at 14mm but was not able to come up with any conclusive examples of this happening at the boundary of IBIS performance (i.e. I couldn't find an example with an obivously blurrier corner than the center).

The errors are important at 14mm in the borders.

I reached the same conclusions as the op. I understand your point and how you model camera shake, but even a slight shake (1 pixel move in the center) will have an important impact in the borders.

I am lazy to make the test (and I don't have 14mm neither FF, only 17mm APS-C, this makes a big difference in fact).

The op has a good point, I had never realised this before !

OP Kenneth Almquist Regular Member • Posts: 162
Re: Tremors are oscillations - not monitonically increasing...
1

kenw wrote:

As to the second conclusion that OIS would be better this overlooks an issue with OIS - it can't compensate for rotation around the optical axis. This can be a significant contribution to total blur from the physical shutter press of the finger causing the camera to rotate around the lens optical axis (or roll axis if you think of the camera as an airframe). This is one reason without IBIS it is recommended to use a delayed shutter to improve hand held shooting.

Yes. IBIS can in principle do a perfect job of compensating for roll, whereas OIS can't address it at all. The CIPA test method only tests for pitch and yaw, claiming (on the bottom of page 37) that, “when the shooting distance is about 20 times the focal length, Yaw and Pitch are dominant, and the other directions can be practically negligible.”  Roll doesn't depend on shooting distance, so the reference to 20 times the focal length is in there only to explain why they don't test for motion in the X, Y, and Z directions.

What's not clear is whether CIPA's comment applies to all photographers or only some.  I haven't ever spotted roll in my pictures, and I've take a lot of pictures at slow shutter speeds with optical stabilization.  At low shutter speeds I use continuous mode (rather than a delayed shutter), and in the past the first and last pictures in the series (the pictures where I pressed and released the shutter button, respectively) were much more likely than the others to have motion blur, but that's because of changes in pitch, not roll.

It does seem plausible that some photographers could introduce roll by pressing the shutter button.  Another possibility is that when pressing the shutter button the photographer simultaneously tightens the other fingers on the same hand (which are holding the grip).  Depending on how the camera is being held, that could introduce a bit of roll.  So the roll correction provided by IBIS but not by OIS could be important to some photographers.

(I read the white paper you linked to but want to think a bit more about the relationship between angle and shutter speed before I comment on it.)

 Kenneth Almquist's gear list:Kenneth Almquist's gear list
Nikon D7200 Fujifilm X-H1 Nikon AF-S DX Nikkor 35mm F1.8G Nikon AF-S DX Nikkor 18-300mm F3.5-6.3G ED VR Fujifilm 16-55mm F2.8R LM WR +2 more
OP Kenneth Almquist Regular Member • Posts: 162
Re: Tremors are oscillations - not monitonically increasing...
1

kenw wrote:

FWIW I was inspired to do some quick tests with a Z7 and 14-30/4 at 14mm but was not able to come up with any conclusive examples of this happening at the boundary of IBIS performance (i.e. I couldn't find an example with an obivously blurrier corner than the center). This was a quick test of only a few shots at each shutter speed with little care in the setup so I don't think there is anything to be drawn from it really.

A test seems like a good idea. I tried some indoor shots but the only example I found was so subtle it might have been caused by motion in the Z direction. So I decided to try photographing a 19 foot long string of lights, which means a shooting distance of 13 or 14 feet. The widest lens I have is the 16-55mm f/2.8, so I was shooting at approximately 24mm full frame equivalent.

Here are crops from near the top, near the center, and near the bottom:

To interpret these crops, you should know that depending on the orientation, these bulbs can look like two lights:  A bright light at the base of the bulb and a dimmer light at the tip.

Finally, to verify that we are seeing motion blur rather than coma, here is a shot of the string with relatively little motion blur:

 Kenneth Almquist's gear list:Kenneth Almquist's gear list
Nikon D7200 Fujifilm X-H1 Nikon AF-S DX Nikkor 35mm F1.8G Nikon AF-S DX Nikkor 18-300mm F3.5-6.3G ED VR Fujifilm 16-55mm F2.8R LM WR +2 more
kenw
kenw Veteran Member • Posts: 6,932
Re: Tremors are oscillations - not monitonically increasing...
1

Kenneth Almquist wrote:

kenw wrote:

FWIW I was inspired to do some quick tests with a Z7 and 14-30/4 at 14mm but was not able to come up with any conclusive examples of this happening at the boundary of IBIS performance (i.e. I couldn't find an example with an obivously blurrier corner than the center). This was a quick test of only a few shots at each shutter speed with little care in the setup so I don't think there is anything to be drawn from it really.

A test seems like a good idea. I tried some indoor shots but the only example I found was so subtle it might have been caused by motion in the Z direction. So I decided to try photographing a 19 foot long string of lights, which means a shooting distance of 13 or 14 feet. The widest lens I have is the 16-55mm f/2.8, so I was shooting at approximately 24mm full frame equivalent.

Here are crops from near the top, near the center, and near the bottom:

To interpret these crops, you should know that depending on the orientation, these bulbs can look like two lights: A bright light at the base of the bulb and a dimmer light at the tip.

Finally, to verify that we are seeing motion blur rather than coma, here is a shot of the string with relatively little motion blur:

Excellent test! Yes I think that seems to show pretty convincing evidence that the effect you calculate does have real world impact even at modestly wide angles for actual hand tremors. You are only at 1/4 second here, so just a bit over two stops beyond 1/f (where f is FF equivalent), and already there is a clearly noticeable effect.

It will probably be some days before I can carve out time to do a similar test at 14mm on FF but your setup here is brilliant and I'll likely do the same. It would seem based on your test results at 24mm equivalent and the math we should see some really obvious effects in a 14mm test.

This is a really interesting result with implications for a common IBIS use case of doing WA and UWA interior shots inside buildings that don't allow for tripods. It implies that for handheld indoor architecture better results would be had without IBIS but instead cranking the ISO to get a handheld shutter speed and then taking multiple exposures that you average in post processing to reduce noise.

Once again, great post!

-- hide signature --

Ken W
See profile for equipment list

 kenw's gear list:kenw's gear list
Panasonic Lumix DMC-GM1 Nikon Z7 Panasonic Leica DG Summilux 15mm F1.7 ASPH Nikon Z 14-30mm F4 Nikon Z 24-200mm F4-6.3 VR +42 more
(unknown member) Veteran Member • Posts: 5,193
percentage corrected function of focal
4

Hello,

I have continued my work.

Here is the equation I had explained in a previous post, it is the new position of point at distance x from the center with rotation t (t is not really the angle, see previous post).

What is interesting is the derivative to see how much it moves for small rotations (dt)

The equation is quite simple...

IBIS correct with a slope = 1 (x=0) if we consider IBIS corrects exactly the center.

But the point at position x moves with a slope = 1 +( x/f)^2

The percentage of the move that is corrected is simply the ratio of both

1 indicates that 100% of the rotation move is corrected.

It gives this graph (for FF, x = 36/2 mm)) :

Here ar some values:

  • f = 35mm, there is still 80% of the rotation corrected
  • f = 18mm, only 50% corrected
  • f = 14mm 38% corrected
  • f = 9mm 20% corrected

This graph can be interpreted in 2 ways, either the correction accros the image from center to border, either the correction of the border for different focal length. This graph shows absolutely everything.

J A C S
J A C S Forum Pro • Posts: 20,303
Re: Theoretical limits of IBIS away from the center of an image

Kenneth Almquist wrote:

IBIS attempts to correct for camera motion by moving the sensor. This cannot do a perfect job of correcting for camera motion.

Just as a general remark, without a quantification, you cannot correct 6 degrees of freedom, well projected to 2D, with 2 degrees of freedom. Aside from macro, moving forward and back has a negligible effect but the other 5 degrees are still there.

About the rotation you talk - everyone who has taken photos of boxy buildings (which means everyone) knows that the box never quite looks as a rectangle. Even if we did when the sensor is strictly parallel to the plane of he rectangular wall, not being parallel makes it look like a non-rectangular figure; not even a trapezoid in general. You cannot align such a figure with a rectangle just with a translation. Now imagine a grid in the physical rectangle and how it projects under non-perpendicular projection - here is your distortion.

Here is a diagram I quickly googled:

https://www.researchgate.net/figure/Distortion-grid-on-the-camera-plane-and-trapezoidal-distortion-effect-on-eye-image_fig5_228967208

spider-mario
spider-mario Contributing Member • Posts: 989
Re: Theoretical limits of IBIS away from the center of an image
1

Christof21 wrote:

Another important point is to know when IBIS is really in action. Does it start when you press rhe shutter release, which means it is centered ? Or is it always in action (because you may need it during composition) ? This is a VERY important point.

On Olympus cameras, you can choose, at least in burst mode. See the manual of the E-M1 III, page 186:

Image Stabilizer

Choose whether the camera prioritizes frame rate or image
stabilization during burst photography.

  • [Fps Priority]: Shooting speed gets priority over image
    stabilization. The sensor will not be reset to the center
    during sequential shooting.
  • [IS Priority]: Image stabilization gets priority over shooting
    speed. The sensor will be reset to the center per frame of
    sequential shooting. The shooting speed will drop slightly.

I have actually found “Fps Priority” quite useful for bracketed bursts (for merging as an HDR image), because it means that the frames are pretty much perfectly aligned right out of camera.

 spider-mario's gear list:spider-mario's gear list
Canon G1 X III Olympus TG-6 Olympus E-M1 II Olympus E-M1 III Canon EOS R6 +14 more
kenw
kenw Veteran Member • Posts: 6,932
A test, conclusive that this does matter - a lot...
5

The main things I wanted to accomplish were:

  1. Show in an absolute sense how hand tremor wanders at relevant exposure times
  2. Illustrate the magnitude of the effect on corners
  3. Illustrate this is worse in a relative sense the more wide angle we go

I think the results are pretty clear.

Test setup was a precision pinhole used for telescope collimation testing sitting across the room.  Shot with a Nikon Z7 and the Z14-30/4S.  I hand held standing with a fairly standard bracing stance and arm positions.  I shot either with the pinhole in the center of the frame or in the upper left corner.  Obviously my holding position was thus a bit different between the center and corner cases but the results below are way beyond any possible variation just due to holding the camera at a different angle.  If anything when the pinhole was in the upper left corner my stance and brace felt more comfortable and more stable as the camera was pointed down.

I did this for 14mm and 30mm running from 1/30th to 1 second in half stops.  Only posting the more interesting results here.

I wanted to show the results quickly so these crops are arbitrarily resampled by Lightroom in its N-shot comparison display but it is the same resampling in all cases.  Each square is 583x583 pixels in the original images.

Obviously view original sizes to see anything useful!

First up is 30mm at 1 second.

Top row is the center with IBIS off.  This shows just how much I wander hand holding.  Clearly magnitudes more than sufficient to cause issues based on the models presented in this thread.

Middle row is center with IBIS on.  IBIS is doing a pretty nice job here but clearly we are right at the limit for 30mm with many showing some clear blur.

Bottom row is upper left corner with IBIS on.  Yeah that's a really clear effect.  What is also neat is how the motion error is mostly only on the radial axis of the image as the theory would predict.

30mm 1 sec

Next up is 14mm at 1sec. I did not repeat the IBIS off shots here since they would be just about half the magnitude of the 30mm case of course.

Top row is center with IBIS on and bottom row is upper left with IBIS on.  Wow is this an absolute mess for sure!

14mm 1 sec

And now for comparison is 30mm at 0.5 sec.  This round has about the same effectiveness in the center as 14mm at 1.0 sec as we'd expect.  IBIS is following the "n stops improvement" figure of merit as we expect.  But the corners are actually much better than in the 14mm at 1.0 sec case.  So this shows quite well that indeed as we go to UWA the IBIS performance in the sense of "n stops improvement" is more limited or degraded in the corners the wider we go.

30mm 0.5 sec

That's it for tonight.  Next I'm going to go back through these shots and take a few more to establish what the rough improvement is for IBIS at the two focal lengths both in the center and corners.  Just how many fewer "stops of improvement" do I get at 14mm and 30mm in the corners compared to the center?

Kudos again to the OP on this!

-- hide signature --

Ken W
See profile for equipment list

 kenw's gear list:kenw's gear list
Panasonic Lumix DMC-GM1 Nikon Z7 Panasonic Leica DG Summilux 15mm F1.7 ASPH Nikon Z 14-30mm F4 Nikon Z 24-200mm F4-6.3 VR +42 more
(unknown member) Veteran Member • Posts: 5,193
Re: A test, conclusive that this does matter - a lot...

kenw wrote:

The main things I wanted to accomplish were:

  1. Show in an absolute sense how hand tremor wanders at relevant exposure times
  2. Illustrate the magnitude of the effect on corners
  3. Illustrate this is worse in a relative sense the more wide angle we go

I think the results are pretty clear.

Test setup was a precision pinhole used for telescope collimation testing sitting across the room. Shot with a Nikon Z7 and the Z14-30/4S. I hand held standing with a fairly standard bracing stance and arm positions. I shot either with the pinhole in the center of the frame or in the upper left corner. Obviously my holding position was thus a bit different between the center and corner cases but the results below are way beyond any possible variation just due to holding the camera at a different angle. If anything when the pinhole was in the upper left corner my stance and brace felt more comfortable and more stable as the camera was pointed down.

I did this for 14mm and 30mm running from 1/30th to 1 second in half stops. Only posting the more interesting results here.

I wanted to show the results quickly so these crops are arbitrarily resampled by Lightroom in its N-shot comparison display but it is the same resampling in all cases. Each square is 583x583 pixels in the original images.

Obviously view original sizes to see anything useful!

First up is 30mm at 1 second.

Top row is the center with IBIS off. This shows just how much I wander hand holding. Clearly magnitudes more than sufficient to cause issues based on the models presented in this thread.

Middle row is center with IBIS on. IBIS is doing a pretty nice job here but clearly we are right at the limit for 30mm with many showing some clear blur.

Bottom row is upper left corner with IBIS on. Yeah that's a really clear effect. What is also neat is how the motion error is mostly only on the radial axis of the image as the theory would predict.

30mm 1 sec

Next up is 14mm at 1sec. I did not repeat the IBIS off shots here since they would be just about half the magnitude of the 30mm case of course.

Top row is center with IBIS on and bottom row is upper left with IBIS on. Wow is this an absolute mess for sure!

14mm 1 sec

And now for comparison is 30mm at 0.5 sec. This round has about the same effectiveness in the center as 14mm at 1.0 sec as we'd expect. IBIS is following the "n stops improvement" figure of merit as we expect. But the corners are actually much better than in the 14mm at 1.0 sec case. So this shows quite well that indeed as we go to UWA the IBIS performance in the sense of "n stops improvement" is more limited or degraded in the corners the wider we go.

30mm 0.5 sec

That's it for tonight. Next I'm going to go back through these shots and take a few more to establish what the rough improvement is for IBIS at the two focal lengths both in the center and corners. Just how many fewer "stops of improvement" do I get at 14mm and 30mm in the corners compared to the center?

Kudos again to the OP on this!

Thank you for this, interesting !

I know it takes very long to do it, so this is very much appreciated.

I wanted in a first step to try to align 2 shots, no camera shake (on tripod) with the only difference that the camera is very slightly rotated on the second shot. I will align the center ofc but I want to see what happens in the borders/corners.

What I wanted to do in the First stem

MEDISN
MEDISN Senior Member • Posts: 1,657
Re: Tremors are oscillations - not monitonically increasing...

kenw wrote:

Excellent test! Yes I think that seems to show pretty convincing evidence that the effect you calculate does have real world impact even at modestly wide angles for actual hand tremors. You are only at 1/4 second here, so just a bit over two stops beyond 1/f (where f is FF equivalent), and already there is a clearly noticeable effect.

It will probably be some days before I can carve out time to do a similar test at 14mm on FF but your setup here is brilliant and I'll likely do the same. It would seem based on your test results at 24mm equivalent and the math we should see some really obvious effects in a 14mm test.

This is a really interesting result with implications for a common IBIS use case of doing WA and UWA interior shots inside buildings that don't allow for tripods. It implies that for handheld indoor architecture better results would be had without IBIS but instead cranking the ISO to get a handheld shutter speed and then taking multiple exposures that you average in post processing to reduce noise.

Would be similar approach to Olympus hand-held high-res no?  Set a shutter speed at 1/f and ISO at whatever it needs to be.  Technically IBIS will still be on for the brief moment during the exposure but resets between each of the 16 frames then aligns and stacks to reduce noise.  ISO 6400 looks like ISO 400 or thereabouts.

Good insights in this thread - confirms my experience that UWA performance is lower than standard focal lengths.  But then my UWA use tends to be landscape/seascapes to detail in the corners is probably less noticeable than with something like architecture.

J A C S
J A C S Forum Pro • Posts: 20,303
Re: Tremors are oscillations - not monitonically increasing...

MEDISN wrote:

kenw wrote:

Excellent test! Yes I think that seems to show pretty convincing evidence that the effect you calculate does have real world impact even at modestly wide angles for actual hand tremors. You are only at 1/4 second here, so just a bit over two stops beyond 1/f (where f is FF equivalent), and already there is a clearly noticeable effect.

It will probably be some days before I can carve out time to do a similar test at 14mm on FF but your setup here is brilliant and I'll likely do the same. It would seem based on your test results at 24mm equivalent and the math we should see some really obvious effects in a 14mm test.

This is a really interesting result with implications for a common IBIS use case of doing WA and UWA interior shots inside buildings that don't allow for tripods. It implies that for handheld indoor architecture better results would be had without IBIS but instead cranking the ISO to get a handheld shutter speed and then taking multiple exposures that you average in post processing to reduce noise.

This will introduce more or less the same error; but you can reject the worst frames if you have the tools to analyze the alignment. Another option would be software which can correct for such errors by doing geometric corrections at the expense of some resolution loss. I believe some focus stacking software does this (enfuse?).

Would be similar approach to Olympus hand-held high-res no? Set a shutter speed at 1/f and ISO at whatever it needs to be. Technically IBIS will still be on for the brief moment during the exposure but resets between each of the 16 frames then aligns and stacks to reduce noise. ISO 6400 looks like ISO 400 or thereabouts.

Good insights in this thread - confirms my experience that UWA performance is lower than standard focal lengths. But then my UWA use tends to be landscape/seascapes to detail in the corners is probably less noticeable than with something like architecture.

spider-mario
spider-mario Contributing Member • Posts: 989
Re: Tremors are oscillations - not monitonically increasing...

J A C S wrote:

kenw wrote:

This is a really interesting result with implications for a common IBIS use case of doing WA and UWA interior shots inside buildings that don't allow for tripods. It implies that for handheld indoor architecture better results would be had without IBIS but instead cranking the ISO to get a handheld shutter speed and then taking multiple exposures that you average in post processing to reduce noise.

This will introduce more or less the same error; but you can reject the worst frames if you have the tools to analyze the alignment. Another option would be software which can correct for such errors by doing geometric corrections at the expense of some resolution loss. I believe some focus stacking software does this (enfuse?).

I don’t think enfuse itself does that but the suite that it’s part of (Hugin/PanoTools) indeed can. Hugin would find a best fit for the distortion parameters, and then nona, the remapper, would apply them.

 spider-mario's gear list:spider-mario's gear list
Canon G1 X III Olympus TG-6 Olympus E-M1 II Olympus E-M1 III Canon EOS R6 +14 more
kenw
kenw Veteran Member • Posts: 6,932
Re: Tremors are oscillations - not monitonically increasing...
1

spider-mario wrote:

J A C S wrote:

kenw wrote:

This is a really interesting result with implications for a common IBIS use case of doing WA and UWA interior shots inside buildings that don't allow for tripods. It implies that for handheld indoor architecture better results would be had without IBIS but instead cranking the ISO to get a handheld shutter speed and then taking multiple exposures that you average in post processing to reduce noise.

This will introduce more or less the same error; but you can reject the worst frames if you have the tools to analyze the alignment. Another option would be software which can correct for such errors by doing geometric corrections at the expense of some resolution loss. I believe some focus stacking software does this (enfuse?).

I don’t think enfuse itself does that but the suite that it’s part of (Hugin/PanoTools) indeed can. Hugin would find a best fit for the distortion parameters, and then nona, the remapper, would apply them.

Yes, and I believe that Photoshop's Auto Align also does this as long as you select an appropriate projection mode. And most any decent HDR stacking process as well.

I probably should have mentioned in my post I was assuming the stack was done with more than just naive translation and rotation. That would of course be no different than what IBIS does. It is just such a commonly implemented thing to do something smarter than that it didn't occur to me to call it out.

That said, worth doing a test/demonstration of that with WA or UWA to illustrate the degree to which it improves or doesn't improve the situation.

-- hide signature --

Ken W
See profile for equipment list

 kenw's gear list:kenw's gear list
Panasonic Lumix DMC-GM1 Nikon Z7 Panasonic Leica DG Summilux 15mm F1.7 ASPH Nikon Z 14-30mm F4 Nikon Z 24-200mm F4-6.3 VR +42 more
MEDISN
MEDISN Senior Member • Posts: 1,657
Long Exposure vs In Camera Stacking

I don't have Christmas lights handy in the studio but I photographed an ISO 12233 chart on a 55" LED with my widest lens.

I took shots below handheld free-standing.

  1. A no-IBIS frame at 1/30s, ISO 1600 as a reference 
  2. A best I could achieve with IBIS - 4s at ISO 100 (required a 3-stop ND)
  3. A hand-held high res shot at 1/30s, ISO 1600 just like #1 above.

If you're not familiar with hand-held high res, it's an in-camera stacking mode that shoots 16 frames (20MP) with electronic shutter then aligns and stacks them into a 50MP raw (and/or JPG).

Sharpening left at PS defaults (Amount 40, Radius 1, Detail 25, Mask 0) and Noise reduction (luminance 0 and color set to 25).

100% crops of the bottom right corner

Looking at the "16:9" in the corner of each image, the 1/30s, no IBIS shot is certainly sharper.  The 4s shot is noticeably less noisy but not as sharp.  The 50MP hand-held high res shot is probably the best compromise between the two.  I would have no qualms using it for a 7mm (14mm eq) shot at f/8 (f/16 eq) like this.  Diffraction already kicking in but not terrible result.

spider-mario
spider-mario Contributing Member • Posts: 989
Re: Long Exposure vs In Camera Stacking

MEDISN wrote:

I don't have Christmas lights handy in the studio but I photographed an ISO 12233 chart on a 55" LED with my widest lens.

I took shots below handheld free-standing.

  1. A no-IBIS frame at 1/30s, ISO 1600 as a reference
  2. A best I could achieve with IBIS - 4s at ISO 100 (required a 3-stop ND)
  3. A hand-held high res shot at 1/30s, ISO 1600 just like #1 above.

(As an aside, given that 3. is the result of combining 16 shots, in principle, it should be labeled as ISO 100.)

“In the case where the image recorded by the camera is the result of processing that combines multiple captured images, the EI value corresponds to the sum of the exposures for the images that have been combined. In this case, the EI value should be reported along with a notation indicating that multiple images have been combined.

EI values should be reported using exposure index metadata in the image file header. In the case where multiple captured images have been combined, the number of captured images that were combined and the EI values for the individual captured images should also be reported using metadata.”

 spider-mario's gear list:spider-mario's gear list
Canon G1 X III Olympus TG-6 Olympus E-M1 II Olympus E-M1 III Canon EOS R6 +14 more
ProfHankD
ProfHankD Veteran Member • Posts: 8,727
Re: percentage corrected function of focal

Christof21 wrote:

It gives this graph (for FF, x = 36/2 mm)) :

Looks right to me -- nice analysis!

Of course, the "stretching" of objects off-axis and common levels of field curvature also make the additional blur less obvious for many, but not all, ultrawide scenes. Various fisheye projections and other oddities (e.g., anamorphic lenses) will also be affected in slightly different ways.

IBIS definitely does help reduce the blur CoC even with the widest ultrawides, but the improvement is actually modest with nearly any lens. There's a lot wrong with how IBIS is measured and implemented. The CIPA tests are grossly flawed, only measuring pitch and yaw response (roll is also significant off axis) for a fixed waveform and not accounting for any camera-generated shake or grip issues -- I believe they also only consider the center of the frame. IBIS is generally controlled as a reactive system, rather than predictive, so there's actually a time delay with overshoots and all that fun too. Of course, shake is also a stochastic thing, so results inherently vary quite a bit even from frame to frame. There's also the issue that IBIS generally needs to know the lens focus distance, which isn't always available -- e.g., when shooting with a manual lens. Heck, even pincushion/barrel distortion can mess-up the correction.

To put it bluntly, the best way to stabilize a capture is to stabilize the entire camera. I'm partial to bolting cameras to concrete blocks.    Eventually, the combination of new sensor tech and computational processing will do better than IBIS -- even the computational multi-shot anti-blur in the original NEX-5 did pretty well. I think people in general need to accept that IBIS is generally useful, yet heavily flawed, but is probably the best answer at this point in time.

 ProfHankD's gear list:ProfHankD's gear list
Canon PowerShot SX530 Olympus TG-860 Sony a7R II Canon EOS 5D Mark IV Sony a6500 +32 more
Keyboard shortcuts:
FForum MMy threads