Flat pan or circular pan? #capturing #howto #alignment


David B
 

 I have used two methods of camera positioning for the stereo images.  One is to choose a focal point and point the camera straight at that point for each of the two images (camera moves on a radius around the focal point).  The other is to flat pan the camera so that the center of the image is offset between the left and right.

I am wondering which is better.

It seems for the best view when focused on a single point would be the circular motion because it is more like the natural geometry of the human eyes.  Where this may be worse is when you shift your attention through the image on locations that are not a this focal point.  Then it can be less natural?

Maybe for best viewing of other points throughout the image a flat pan is better?

It seems the closer the object the more effect there is for these choices.

How do you do it and what works best?  When doing close up shots what do you do?


 

I have used two methods of camera positioning for the stereo images.
Ah, yes. The age old question -- to converge or not converge.

I'll stand back and let the folks who have strong technical and
philosophical arguments for and against argue the fine points.

For general 3D photography, my own take is simple. I think of the 3D
camera like a pair of magic glasses. The goal is to capture what you
would see if you could look through those glasses at that time and
place. Thus the converging should be done by the viewer's own eyes
looking at the captured image, just as you would looking through the
magic glasses, rather than converge the lenses of the glasses. But
again, this is a very simplistic view of the problem and 3D macro
photography adds its own complications and limitations. But even
there, I personally use special closeup/macro lenses that keep the
optics parallel, not converged.

...BC


Bob Aldridge
 

You are right about toeing in causing difficult in viewing different distances. And it is a subject that has been debated for many years. A Stereoscopic Society Bulletin from 1964 has an article on the subject from "The Professor" (Bill Dalgoutte) referring back to the British Journal of Photography in 1904 which said that there was no advantage in toeing in. He then alludes to "a popular post-war textbook" that claims that toeing in is natural, because our eyes do it...

The Professor then goes on for the rest of the article explaining why toeing in will render objects at different distances from the one that is toed in to will be progressively more difficult to view...

Another article in a 1969 Bulletin seems to defend E.F. Linssen in his book "Stereo Photography in Practice" (which would appear to be the "textbook" that the Professor mentions) and his view that toeing in is natural... Interestingly, the author of this article is not named, but it seems to be someone with a problem with Bill Dalgoutte - the then Editor of the Bulletin and the "Honorary General Secretary" of The Stereoscopic Society, and also "The Professor" (Collectively known as "The Tolworth Triumvirate" since Dalgoutte lived in a suburb of London called Tolworth...) My guess would be John Singleton, but that is just a guess...

It is interesting that another article in the same Bulletin immediately follows with the title "The mathematical analysis of toe-in distortion" counters it with a demonstration of why toeing in is bad! It is also unattributed, but an allusion to "the books in the library of the Hon Gen Secretary (Bill Dalgoutte, AKA The Professor) would suggest that he was the author...

Anyway, I, personally, think that you can "get away" with mild toeing in, but it is not ideal. And the detrimental effects will be worse the more "depth" there is in the scene. Many of the distortions can be "corrected" in the digital world, of course, which was not the case in the 1960s, but there will always be some negative impact on the final image.

So I start wondering why toeing in would be desirable. For cameras with "normal" lens spacing, taking images of subjects between about 2 metres and infinity, the only benefit is to avoid cropping opposite sides to set the "window", but todays high resolution cameras yield images that reduce the quality impact of that cropping...

But, for close-up photography, it is a different story. If you cannot get the lenses close enough together, it is possible that one camera will not even see the main subject! Here, toeing in MIGHT be an answer. But it will be vital to seriously control the depth in the scene. If you don't you could easily get totally different backgrounds! A better approach might be to use telephoto lenses so the camera is much further from the subject obviating the need to toe-in. But, again, the depth in the scene must be controlled or the effect that Arthur Girling called "Reverse Frustrum Distortion" will become evident!

As with EVERYTHING in stereoscopy, there are compromises. And that is why it is so fascinating.

Just for the record, I believe that the ideal is to keep both sensors in the same plane. and that means that toeing in should be avoided without a very good reason to do it... It should not be used to set the window, for instance.

Bob Aldridge



On 29/01/2021 01:07, David B via groups.io wrote:
 I have used two methods of camera positioning for the stereo images.  One is to choose a focal point and point the camera straight at that point for each of the two images (camera moves on a radius around the focal point).  The other is to flat pan the camera so that the center of the image is offset between the left and right.

I am wondering which is better.

It seems for the best view when focused on a single point would be the circular motion because it is more like the natural geometry of the human eyes.  Where this may be worse is when you shift your attention through the image on locations that are not a this focal point.  Then it can be less natural?

Maybe for best viewing of other points throughout the image a flat pan is better?

It seems the closer the object the more effect there is for these choices.

How do you do it and what works best?  When doing close up shots what do you do?


Etienne Monneret (Perso)
 

Le 29/01/2021 à 02:07, David B via groups.io a écrit :
It seems for the best view when focused on a single point would be the circular motion because it is more like the natural geometry of the human eyes.  Where this may be worse is when you shift your attention through the image on locations that are not a this focal point.  Then it can be less natural?
Maybe for best viewing of other points throughout the image a flat pan is better?

When you are shooting with cameras aligned perfectly parallel, the image is a circular scan of the field of view. When viewed at screen, this circular scan is perfectly restored on your screen, and you have to converge to see it exactly like in a natural viewing.

When you are shooting with convergent cameras, when restored at screen, the rotation appears as a wrong viewing angle. Your screen is flat, and you do not have the possibility to restore the original viewing angle of your cameras. At least, you need to get two screens to display with a viewing angle between both to restore the shooting conditions.

Here is a small diagram:

  1. top left: at shooting time, the cameras are well parallel. The green line is what your cameras are shooting, supposed to be displayed later on your screen. Red and blue lines are displaying what is shot for a far and a near object.
  2. top right: you display both pictures aligning the nearest point to fit the same place on your screen, with a zero parallax. All the scene is restored for your eyes exactly at the same place as the shooting time, in a natural viewing angle for the far object. This is the only way you can get the exact viewing effect on your single flat screen as at the shooting time.
  3. bottom left: at shooting time, the cameras are pointing a same place on the scene. The green line is style what your cameras are shooting.
  4. bottom right: you display both pictures, still aligning the nearest point to fit the same place, with a zero parallax. You see that the original far object now have a much larger parallax at screen, even with a high risk of divergence. If you want to restore the original parallax of the far object, you need to put the nearest object in front of your screen, with a strong negative parallax. You simply can't get back the natural viewing angles between both views as at the shooting time, except using 2 screens with an angle between them.



Etienne Monneret (Perso)
 

This diagram is also an occasion to explain the 1/30 cooking rule.

Suppose you  are shooting a mountain, the diagram is like this (and it's easy to understand the contrary if you are shooting a bee), with the shooting time on the left, and the viewing time on the right:



Etienne Monneret (Perso)
 

Remark: I agree that, with my diagram, a 6 cm distance to your screen is a bit short. I should have adapted a bit more. But, the idea is here.
:D

Le 29/01/2021 à 09:22, Etienne Monneret (Perso) a écrit :
This diagram is also an occasion to explain the 1/30 cooking rule.

Suppose you  are shooting a mountain, the diagram is like this (and it's easy to understand the contrary if you are shooting a bee), with the shooting time on the left, and the viewing time on the right:




JackDesBwa|3D
 

I am wondering which is better.

In short: none.

The ideal way is to use shift lenses, but it is expansive unless you shoot inside a 3D modeling software.
Parallel and toed configurations are ways to approximate this ideal one, with their own drawbacks.
Mathematically (and thus digitally), you can transform the two into the ideal situation, with the nuance that the final field of view should be included in the picture of course. (In analog world, reproducing the transformations is possible, but might require complex installation for toed configuration)

In parallel configuration, you have to correct the translation introduced because the lens is not at its ideal position, which is a pretty simple operation. If not corrected, you have (possibly large) window violations because the window is set at the infinity. Of course, you have to remove areas on the sides with this operation.
In toed configuration, you have to correct the perspective introduced because the sensors are not parallel, which is a slightly complex operation that some software do not do well. If not corrected, you have vertical disparities going bigger on the sides (and depending on the distance of the objects) which go against the ability to fuse and against comfort, and a bombing of the depth (sides appear farther than it needs compare to the central part). This operation removes parts of the image too, but it is less obvious.

With perfect correction, capturing in parallel or toed conditions, assuming the "nodal point" of the lenses are at the same place in the scene, you will end with the exact same image.

You can thus choose either method, depending on your ability to do the right correction and the one that is more adapted to your shooting conditions (e.g. as Bob said, it might me hard to shoot close-up with parallel configuration when you have two lenses, but in toed-in configuration without telephoto lens you also have large base and would have to limit the depth).
In practice, there is always a slight misalignment of both kinds so that you often have to correct both effects (in different magnitude) to have a perfect image.

because it is more like the natural geometry of the human eyes

This assertion is false, or at least it forgets to take into account that the photo is displayed on a flat medium.
The drawing of Etienne shows it.

You can also imagine a photo of a flat textured wall. When printed (in anaglyph for example) and placed on the wall, it should look the same.
Thus you have to correct the perspective in toed configuration [otherwise the perspective is applied twice: once with the camera, once with the eyes].

JackDesBwa


Etienne Monneret (Perso)
 

Hummm... I realize that my demonstration is wrong, because my diagram was not rotated in the good way.

Forget it.

JackDesBwa explanations are much more pertinent than mine.

Le 29/01/2021 à 08:43, Etienne Monneret (Perso) a écrit :
Le 29/01/2021 à 02:07, David B via groups.io a écrit :
It seems for the best view when focused on a single point would be the circular motion because it is more like the natural geometry of the human eyes.  Where this may be worse is when you shift your attention through the image on locations that are not a this focal point.  Then it can be less natural?
Maybe for best viewing of other points throughout the image a flat pan is better?

When you are shooting with cameras aligned perfectly parallel, the image is a circular scan of the field of view. When viewed at screen, this circular scan is perfectly restored on your screen, and you have to converge to see it exactly like in a natural viewing.

When you are shooting with convergent cameras, when restored at screen, the rotation appears as a wrong viewing angle. Your screen is flat, and you do not have the possibility to restore the original viewing angle of your cameras. At least, you need to get two screens to display with a viewing angle between both to restore the shooting conditions.

Here is a small diagram:

  1. top left: at shooting time, the cameras are well parallel. The green line is what your cameras are shooting, supposed to be displayed later on your screen. Red and blue lines are displaying what is shot for a far and a near object.
  2. top right: you display both pictures aligning the nearest point to fit the same place on your screen, with a zero parallax. All the scene is restored for your eyes exactly at the same place as the shooting time, in a natural viewing angle for the far object. This is the only way you can get the exact viewing effect on your single flat screen as at the shooting time.
  3. bottom left: at shooting time, the cameras are pointing a same place on the scene. The green line is style what your cameras are shooting.
  4. bottom right: you display both pictures, still aligning the nearest point to fit the same place, with a zero parallax. You see that the original far object now have a much larger parallax at screen, even with a high risk of divergence. If you want to restore the original parallax of the far object, you need to put the nearest object in front of your screen, with a strong negative parallax. You simply can't get back the natural viewing angles between both views as at the shooting time, except using 2 screens with an angle between them.




George Themelis
 

To converge or not?

 

As Bob Aldridge said, in close-up photography you might have to converge, in order to maintain any overlap between the R and L views (if you do not converge, the right camera can see a completely different view than the left).

 

I came to 3D via Scanning Electron Microscopy (SEM). At the high magnifications (not excessively high, even at 50-100x) of the SEM, the only way to record a stereo pair is by tilt (tilting of the sample), so pure convergence, no parallel shift.

 

A simple formula for the amount of “image loss” during shift is: M * B

 

Where M is the magnification, and B is the stereo base or amount of shift.

 

The higher the magnification or the larger the shift, the more image is lost and the more convergence becomes a necessity.

 

These days I shoot a lot of close-ups with twin cameras side-by-side and long focal length lenses. The cards are stacked against me (high M, high B). I am forced to converge my cameras or I am in trouble. I have devised a system to converge them smoothly and of course I need to change the convergence depending on the distance to the subject.

 

Another area where one might want to converge are hyperstereos. The magnification is not the issue but the stereo base (B) could be large, so image overlap becomes a problem. Also, for practical reasons, when you use a single camera and shift, after you take the first picture, you tend to center an object while you shift, to maintain alignment, so you are converging the camera during the shots.

 

These are specialized areas of 3D photography. For normal stereo photography, I try keep my cameras parallel.

 

George

 

 

 

 

 

From: Bob Aldridge
Sent: Friday, January 29, 2021 2:43 AM
To: main@photo-3d.groups.io
Subject: Re: [Photo-3d] Flat pan or circular pan? #alignment #capturing #howto

But, for close-up photography, it is a different story. If you cannot get the lenses close enough together, it is possible that one camera will not even see the main subject! Here, toeing in MIGHT be an answer. But it will be vital to seriously control the depth in the scene. If you don't you could easily get totally different backgrounds! A better approach might be to use telephoto lenses so the camera is much further from the subject obviating the need to toe-in. But, again, the depth in the scene must be controlled or the effect that Arthur Girling called "Reverse Frustrum Distortion" will become evident!

As with EVERYTHING in stereoscopy, there are compromises. And that is why it is so fascinating.

Just for the record, I believe that the ideal is to keep both sensors in the same plane. and that means that toeing in should be avoided without a very good reason to do it... It should not be used to set the window, for instance.

Bob Aldridge

 


Depthcam
 

The question has a slightly different answer now then it had in film days.

For geometric correctness, lenses should be parallel.  This means that - should you photograph a brick wall, there will be no trapezoidal discrepancies between the left and right images.  Cropping for the stereo window can be achieved either by shifting the lens or just trimming the sides.

As others have said, there are situation where optical convergence cannot be avoided such as when shooting tele-macro, hyperstereos, etc.  Here it's important to emphasize that the amount of convergence required is very small and can most of the time be compensated in post with programs such as SPM.  The more convergence there is however, the more problematic it becomes - for two reasons: First the opposite keystone distortion increases and becomes more difficult to correct in post and secondly, subjects in the background are no longer aligned because the lenses are looking in different directions.  Therefore when converging, it's best to keep the background very close to the main subject or to make it as plain as possible.  This is especially true when using single lens SBS mirror adapters which change the path of the optical axes drastically to record the same image on the left and right side.

So the best approach is to avoid convergence as much as possible.  However, when you must use it, keep it minimal and keep your background either close or plain.

There are circumstances where convergence is very difficult to avoid.  For example when using multi-cam rigs where the subject needs to remain in the center of the frame.  But with two-cam rigs. it's best to avoid it as much as possible.

Francois


Antonio F.G.
 

To converge or not?

For stereo viewing the answer is clearly NOT. The pair of images seen by the human MUST be equivalent to the ones taken by a pair of cameras perfectly parallel, and with the sensors in the same plane. This is the only way to prevent the dreadful vertical disparity.

And for shooting? Well, you can shoot with any convergence angle that suits for the scene, but then you must apply a perspective transform to get the images that would have been taken with an ideal parallel stereo rig. This is what stereo alignment is about.
Even though you can align ANY convergent/divergent pair, the alignment is easier if the pair is shot approximately parallel. But again, if need be there should not be any problem to converge the shoot.

This animated GIF shows how the images from two wildly divergent cameras are perspective-transformed to produce a perfectly aligned pair:

Regards
     Antonio


Olivier Cahen
 

Convergence in shooting leads to vertical disparity in  the corners of the image, and to a bad représentation of the distance scale.

Best regards, Olivier

Le 30 janv. 2021 à 15:51, Antonio F.G. <afgalaz@...> a écrit :

To converge or not?

For stereo viewing the answer is clearly NOT. The pair of images seen by the human MUST be equivalent to the ones taken by a pair of cameras perfectly parallel, and with the sensors in the same plane. This is the only way to prevent the dreadful vertical disparity.

And for shooting? Well, you can shoot with any convergence angle that suits for the scene, but then you must apply a perspective transform to get the images that would have been taken with an ideal parallel stereo rig. This is what stereo alignment is about.
Even though you can align ANY convergent/divergent pair, the alignment is easier if the pair is shot approximately parallel. But again, if need be there should not be any problem to converge the shoot.

This animated GIF shows how the images from two wildly divergent cameras are perspective-transformed to produce a perfectly aligned pair:
<alineamiento3-Presentación1.gif>
Regards
     Antonio_._,_._,_


John Clement
 

The discussion over how to create stereoscopic pairs the provide comfortable viewing really resolves to “It depends”.  Your eyes are very adaptive instruments.  Take the case of someone with severe astigmatism.  Essentially this produces an anamorphic image with attendant blurring.  Glasses that fix this problem reverse the process and produce much clearer vision.  However when you hold the glasses away from the eyes you can clearly see hos the image is squished in one dimension.  My eyes have astigmatism axes at right angles.   When younger I had 20/10 vision with both eyes because the brain unscrambles this and fixes the blurred detail in each eye.   If I look up through the glasses images look good despite the anamorphic distortion, but if I look up even higher I get split vision.  Take the glasses off and there is still some split vision looking far up because I have accommodated to the glasses.

 

View-Master reels often have bad vertical alignment, but when looking through the viewer the images look good without distress.  Rotations, however, can produce distress and viewing problems.  So why do the VM pictures look good?  The images are seen without other external objects to establish a reference frame.  As a result they eyes naturally align so as to accommodate the disparity.  So stereoscopic images seen through some sort of viewer scope do not need strict alignment.

 

Images seen on an external screen such as a 3D movie or TV had other images nearby, so the disparity is obvious when there is not good vertical alignment.  With a hand held device this can sometimes be accommodated by a slight rotation of the device, but there are still some disparities.  A well darkened theater with huge screen dampens the problem by obscuring nearby details, but a home TV is more problematic.

 

Because of the distortion in human eyes the stuff in the central vision will look good, but there may be split vision on the peripheray.  When you look at things to the side, the eyes reaccommodate.  I am so used to my glasses that I can look to the side and the objects instantly look good, but with them off I get momentary split vision.

 

Si I submit that all stereoscopic images are not perfect representations of what I see (perceive?).  Basically the goal is to create images that look real enough to fool the average, or perhaps my own eye.  Even flat TV or pictures in color cannot give the same sensations of color as real objects, but they can come close.  3 colors cannot stimulate the cones exactly the same way as a given mixture of natural light.  However, the information presented to the brain is actually the difference across boundaries, so 3 colors can fool the brain as long as the difference is correct.  I saw Edwin Land prove this many years ago.  In addition everyone sees colors differently so what one calls green another can call blue or blue-green.  Everyone sees 3D differently.

 

When I look at my VM or Realist 3D images I see depth and rounded objects.  But a friend sees them as cardboard cutouts.  Another friend with poorly aligned eyes does not normally see stereoscopic images.  However, one time when she went to a 3D movie, she did see them and asked her husband if that is what normal people see!  Another friend who has the same problem was amazed by looking through an inherited Realist viewer because he could see 3D.  He became an instant 3D photographer.  There is an artist who had this problem and through extensive training and exercises came to see stereoscopic 3D.  She claimed it improved her artwork.  Perfect 3D for everyone is not possible, but enjoyable 3D is achievable.  Artifice and exaggeration can be used to produce more WOW images that look OK to most viewers.

 

Remember “Perfection is the enemy of the good”.

 

John M. Clement

 

 


David B
 

I appreciate all the detailed discussion here.  Many of mine are shot with two sequential images, one camera, with just an approximate inter-ocular distance horizontal shift.  There will always be some convergence and rotation mismatch to deal with,sometimes it turns out worse than other times.  I've had good luck with long distance hypers for the most part, but less and less as the object gets closer, especially when confounded by background objects that end up poorly represented and distracting.  I haven't really had any issues of loosing too much of the image due to flat plane camera positioning.  It seems the closer I am to the object the less distance between the images I want to use anyway, so that mitigates the problem.

Antonio, I like that animation you made showing the issue.  Are there any easy processing steps you take to find the optimum perspective transformations for a given set?  Or is it as I have been doing, just use shift and skew to play with it until it seems OK?  For some images it is a little (or a lot) difficult to get those perspectives corrected optimally.  Maybe made worse by my hand held shooting methods, which I am looking to improve from this discussion.

Thanks.


Etienne Monneret (Perso)
 

Le 31/01/2021 à 23:25, David B via groups.io a écrit :
Antonio, I like that animation you made showing the issue.  Are there any easy processing steps you take to find the optimum perspective transformations for a given set?  Or is it as I have been doing, just use shift and skew to play with it until it seems OK?  For some images it is a little (or a lot) difficult to get those perspectives corrected optimally.

I'm not sure of what you mean with "shift and skew", but perspective deformations (keystones) you get when cameras are not well parallel are not simply translations, rotations, nor simple vertical or horizontal scaling. They are trapezoidal deformations where the small part of the trapeze is compress in both direction (zoom out), while the large part of the trapeze is expended in both direction (zoom in).

Doing a strong perspective deformation of a regular grid with the 3D tool of Gimp provides you with something like this, where you can see that nearest squares are larger in both directions:

3D-Grid.jpg


You may try to compensate with a post-processing using translation, rotation, horizontal and vertical scaling, or even with a trapezoidal transformation, it could bring you with a very honorable result. But, the stereo won't be completely natural until you really make a complete perspective transformation including its progressive zoom in/out effect.

Antonio is providing with a software that evaluates (human assisted) and process the deformation on both views of a stereo:

http://afgalaz.es/stmani3/en/

https://sourceforge.net/p/stmani3/code/ci/master/tree/

I find his software very impressive and unique. It's too bad that it needs some minimal developer skills to install it. It's a break for a large adoption of it by the stereo enthusiast community.

:)




JackDesBwa|3D
 

Are there any easy processing steps you take to find the optimum perspective transformations for a given set?  Or is it as I have been doing, just use shift and skew to play with it until it seems OK?

As Étienne said, the optimal transformation (perspective correction, that is with trapeze deformation) is slightly complex and hard to do by hand. However, it can be easily modeled with matrix operations (assuming the lens deformations are corrected beforehand) so that software can work with them. The difficulty is to find the right matrices, especially in our classical case where cameras were not calibrated (the calibration works only if the cameras and their focal are fixed), which is usually done by finding a bunch of homologous points and solving an optimization problem to align them as much as possible (with correctly found [and well placed] homologous points, sub-pixel accuracy is achievable).

Antonio is providing with a software that evaluates (human assisted) and process the deformation on both views of a stereo

His software is great for that.
It also has an automated batch mode, though the interface to reach it is a bit complex (several conditions to make it appear in the menu that pops if you click at the right place if I remember well ; refer to the online doc).

It's too bad that it needs some minimal developer skills to install it. It's a break for a large adoption of it by the stereo enthusiast community.

Actually, it is far easier than it was some months ago ;-) but it could probably be improved more.

JackDesBwa


Antonio F.G.
 

On Sun, Jan 31, 2021 at 04:25 PM, David B wrote:
Are there any easy processing steps you take to find the optimum perspective transformations for a given set?  Or is it as I have been doing, just use shift and skew to play with it until it seems OK?  For some images it is a little (or a lot) difficult to get those perspectives corrected optimally.  Maybe made worse by my hand held shooting methods, which I am looking to improve from this discussion.
STMANI3 finds the perspective transformations that minimize the vertical disparity.
It is a successive approximation method that needs to start with an assumption for the perspective parameters. By default it assumes all perspective and rotation angles are zero, and the focal length is the one stated in EXIF data. This works well for well shot stereo pairs, that is, when the cameras are approximately parallel.

When the cameras are far from parallel, the successive approximation will normally not converge, or converge to a wrong result. In this case the user can help the calculation modifying the initial values. This is explained in the "Tips and Tricks" chapter of the user manual: http://afgalaz.es/stmani3/en/angular_alignment.html#tips-and-tricks

Regards
   Antonio


JackDesBwa|3D
 

I just saw this image and I think it illustrates well the effect seen when the sensors are parallel or not: (before correction)

toein.jpg

JackDesBwa