Mail
NEW

Experiences with the FinePix REAL 3D W1

See for information on Fuji W3 here.

Pictures/videos made with FinePix REAL 3D W1

See Flickr and YouTube

Effective size of the CCD in 3D mode (digital zooming)

When one toggles the 2D and 3D mode switch on the REAL 3D W1, one sees a change in VoF. In the 3D and ADV.2D modes there is smaller VoF than for 2D and ADV.3D modes. This is due to in-build shifting, cropping and then resizing done in the camera.

A test have been done to determine the cropping:

put the camera at a fixed location (tripod or just keep it still during the rest of the steps)
put the camera in 2D mode (by pressing the button 2D)
make a picture (this becomes a mono jpg file)
put the camera in 3D mode (by pressing the button 3D)
make a picture (this becomes a stereo mpo file)
split the mpo file (e.g. using FinePixViewer or SPM) into left and right jpg file. Here is the left jg file:
now compare the mono jpg file with the left jpg file

One sees a difference in FoV (comparing the EXIF data in each files one sees no difference in the recoding parameters).
If one loads the above pair into SPM (Easy Adjustment function) and one decreases the right image size (originally the left side of the stereo mpo file) to ~88%; the two pictures map. Also a small shift is needed fully map the two picutres!
So it looks that internal the camera, the stereo pair is cropped (with ~88%) and then resized to the normal picture size of 10 MP). Due to the enlarging of the picture the perceived noise also increases (with the same amount).
In the above sample an F_c = 6.3 mm (minimum zoom) has been used, the same can be done for F_c = 18.9 mm (maximum zoom):

If one loads the above pair into SPM (Easy Adjustment function) and one decreases the right image size (originally the left side of the stereo mpo file) to ~70%; the two pictures map.

So the effective size of the CCD has been cropped in the 3D (and ADV.2D) mode with ~88% (for F_c = 6.3 mm) and ~70% (for F_c = 18.9mm). This reduces the number of effective pixels from 10 MP in 2D mode to ~7.8MP (for F_c = 6.3 mm) and ~4.9 MP (for F_c = 18.9mm) in 3D mode.
These values also map the change of focal length as described in 'Zoom' section of the manual (on page 105).

The behavior of the cropping (black) and thus the effective CCD size (pink) depends on the focal length as can be seen here:

A PSP script is avaibale to compensate for the in-build shifting, cropping and barrel/pincushion distortion (all under your own risk). Please ask me.

Conclusion

The above relativates the 10 MP of the camera in 3D and ADV.2D mode: it looks to be between an effective 4.9 MP and 7.8 MP depending for the focal length. In 2D and ADV.3D mode the effective size of 10 MP looks to be correct.
Remember that the text Twin CCD 10.0 mega pixels on the camera is factual.

As the effective 3D picture is enlarged to 10 MP (3,648 x 2,736) again, the experienced pixelation noise also increases.

Cropping is a logical action; did we not do that in former times with film;-). In the Fuji, as it is digital, they can do after a crop a resize, which was not do-able in film times;-). But it looks Fuji is not doing an optimum crop, see next section.

Proposal for Fuji: Don't resize the cropped file, so no digital zoom (to reduce noise).

Can this cropping be optimized?

To determine the difference between cropping by the Fuji (in 3D and ADV.2D mode) and using uncropped pair (in 2D and ADV.3D mode) plus SPM, the following steps were done:

stereo picture was made in 3D mode (with focal length of 6.3 mm and 18.9 mm)
the 3D MPO picture was resized back to their original dimensions (resized to resp. 88% and 70%)
using cha-cha method: left and right pictures were made in 2D mode (with focal length of 6.3 mm and 18.9 mm)
using SPM: align (with same stereo window as the picture made in step 2) the left and right pictures (with focal length of 6.3 mm and 18.9 mm).

When comparing the results of step 2 and step 4, one sees that the size of the objects/forms are the same (as expected), but the effective area of the pictures made in 3D and ADV.2D mode (step 2) is smaller (lost the black area) than when using the 2D and ADV.3D mode.
For these specific examples, it looks that Fuji overcrops the pictures:

For focal length 6.3 mm:

When using 2D mode the effective CCD size if 9.6 MP
When using 3D mode the effective CCD size if 7.8 MP

For focal length 18.9 mm:

When using 2D mode the effective CCD size if 8.7 MP
When using 3D mode the effective CCD size if 4.9 MP

Proposal for Fuji: Fuji to adds an extra 3DREC feature: beside MPO and MPO+JPEG also have MPO+UNCROPPED.

Proposal for Fuji: Fuji could optimized the cropping better, let me know if advice is needed.

Proposal for Fuji: Why are the ADV.2D pictures also cropped and resized? Please keep all 2D pictures the full size.

2D image noise/sharpness

I compared the image noise/sharpness in Fuji W1 picture (2D mode) with the Canon G11. I used the following work flow:

Both cameras in a size mapping CCD size, best quality and best color system (although color system is not that important, if looking at noise).
Make sure that they have the same ISO value
There are two methods for comparing when CCD resolutions and/or sizes are different::

Compare same viewing size of picture (looking at quality with fixed size/viewing distance in mind)
Have 35mm equivalent focal lengths and apertures of both cameras the same (in that case one needs to resize one of the resulting pictures to get the same width dimension on screen) or
Compare pictures have the same pixel width (so this one is looking more at perceived quality with enlarging in mind)
Put focal lengths and apertures depending on CCD resolution of cameras

Use a focal length for each camera so that the number of pixels for an object in both pictures is the same. The relation one needs to keep in mind is:
F₁=F₂*CCD₂/CCD₁
with:
F₁=focus length of camera 1 [mm]
CCD₁=CCD resolution of camera 1 [pixels/inch]
F₂=focus length of camera 2 [mm]
CCD₂=CCD resolution of camera 2 [pixels/inch]
Keep the Circle of Confusion diameter (CoC) the same, so the aperture of the cameras should be related as follows:
f#₁=f#₂*F₂/F₁
f#₁=f#₂*CCD₁/CCD₂
with:
f#₁=aperture of camera 1 [-]
F₁=focus length of camera 1 [mm]
CCD₁=CCD resolution of camera 1 [pixels/inch]
f#₂=aperture of camera 2 [-]
F₂=focus length of camera 2 [mm]
CCD₂=CCD resolution of camera 2 [pixels/inch]

the dpreview method

Make sure that the scene is quite flat in depth, so that focusing of the two cameras is happening in the best way (and one camera does not focus on something else and thus blurring the to be compared part). Also have a stable scene (so not influenced by wind).
Make comparable pictures at closely the same time, but not in a sunny environment (the darker regions give the best impression of noise). Have the object around which the noise will be compared in the middle of the pictures (to reduce differences in barrel/pincushion distortion).
I used the REAL W1 in 2D mode. Fuji W1 in 3D mode (due to the build-in digital zoom) will have by definition more noise than its 2D mode (due to its inherit digital zoom). If one resizes the Fuji W1 3D mode picture to is effective size; no real noise difference is seen with the Fuji 2D mode picture.
If one camera can provide RAW, convert the RAW to TIF and also convert Fuji W1 jpg into TIF. Don't do any optimizing, shifting, changes, etc. to the pictures (so keep them for instance also in side by side).
For method 1, a different size % for one picture is needed to be able comparing (remember this will add inter/extrapolation errors)
For method 2: compare both pictures at 100% size, so no on screen resampling happens.

I made screen grabs in PSP when looking at the Fuji W1 and Canon G11 pictures (100%). Below are the screen grabs for 100, 400 and 800 ISO:

100 ISO and both pictures at 100%

400 ISO

400 ISO and both pictures at 100%

800 ISO

800 ISO and both pictures at 100%

From the above pictures it looks that the Canon G11 has somewhat more noise/less sharpness, but it provides somewhat better definited white areas. So looking at that, they migth have comparable noise/sharpness levels. Perhaps I need to use a better measure of picture noise/sharpness.
Let me know if you agree or that I made some procedural error, if there are good measure to determine noise/sharpness levels, etc (in which case I can adjust my page). So any feedback is very welcome.

In-build shifting of the pictures

Beside the above cropping, the pictures are shifted in the camera. Depending on the focal length the shift in the left picture changes in my camera (other cameras have different values: Dmitry Semyonov, pers. comm. [2010]) according to the following curve:

In-build shift in left picture

The in-build shift curve for the right picture can't be determined directly as no clean (un-processed) 2D picture can be retrieved from the right lens. But one can reverse engineered it. As the physical convergence angle is 2.2 deg for the right lens, the in-build shift becomes:

Shift in right picture

If a physical onvergence angle of 1.5 deg would have been implemented, the in-build shift needed would be almost constant:

Physical convergence angle

The observation of Arturo Canalda (fuji3d-yahoo [2009]) migth be important for the direction of the lenses:

Hold the camera as if you are going to take a picture and look from the upper side both lenses. In this position you can see the glass that covers the right lens is in a slight different plane that the left one.

And see this blue tag print, which th ewebmaster made of the outside of the right lens:

Slanted surface of the REAL 3D W1 rigth lens

If one measures this slanted surface, one gets an angle of ~2.3 degrees (measured in PSP), so quite close to the 2.2 degrees mentioned as convergenced angle in the EXIF data.

Looking at the EXIF data the physical convergence angle (Ferwerda [1990], chapter 9) the following can be deducted:

The left lens has no physical convergence angle according to EXIF data; this one is also the one used for 2D shots, so that sounds good.
The right lens has a physical convergence angle of 2.2 deg according to the EXIF data; the right lens is positoned under an angle (pers. comm. C. van Ekeren [2010])
This means that some key-stoning happens on the right lens (this has consequences for the use of the Auto alignment function in SPM)

The physical convergence angle of the left and right lens is fixed, as no angular movement of the lens can be seen when changing the focal length.
A more optimum physical convergence angle could have been chosen perhaps, see here.

Proposal for Fuji: Why is a physical convergence angle used?.

Apparent convergence point or stereo window position

The stereo window position in 3D mode (or what the W1 calles: apparent convergence point) has been determined in two ways.

By using infinite points

By using the infinite points in pictures the combined apparent convergence angle is determined:

The convergence point at 2 m (F_c=6.3 mm; see the manual, page 105) gives a combined apparent convergence angle of 2.2 deg, as also seen in EXIF data.
The 2.2 deg makes a 140 pixel deviation for infinite points in the two pictures
In an actual pair you see a deviation of 177 pixels for infinite points. Compensation this for the resizing (at F_c=6.3: 88%) makes an effective deviation of 177*.88=157 pixels. So close-ish to 140 pixels.
Making an actual picture at F_c=18.9 mm, the deviation of the infinite points is here 208 pixels.
With a resizing of 70%, this gives an effective deviation of 146 pixels.
And this is a combined apparent convergence angle of atan(146/595.2/18.9) = 0.74 degrees, which is close to a convergence distance of 6 m..
This combined apparent convergence angle at F_c=18.9 mm is close-ish to what the manual says (atan(77/6500) = 0.68 degrees).

So the combined apparent convergence angle looks to change as the manual implies on page 105.

By using nearby plane

The apparent convergence angle for each lens using a nearby plane was done in the following way:

put the W1 camera 3300 mm (walldistance) from a wall that has a ruler (or nice wallpaper that can act as a ruler)
lensseperation of W1 is 77 mm
make stereo pictures with different focal lengths
measured in the right and left picture where the middle (apparent optical axis) of the picture is, using the ruler on the opposite wall. This gives rightlens and leftlens
determined what the apparent convergence angle is
(= 2 * atan(((rightlens - leftlens)/2+lensseperation/2)/walldistance))
both the right and left lens seem to be apparently converging the same amount; which can be more or less explained by the shift.

According to the above measurements, the combined apparent convergence angle is between 2.2 degrees (equivalent to 2 m stereo window position) and 0.78 degrees (equivalent to 5.7 m stereo window position); different but close(?) to what is in the manual.

Conclusion

It is expected that no physical movement of the CCD is needed to explain the W1 behavoir.
This change of combined apparent convergence angle is not found in EXIF data; which holds the physical convergence angle.

Looking at the camera standard (more particular the Explanation section, page 20, etc.); Fuji follows the ideas about convergence adjustment and horizontal shifting as described in that document.

Proposal for Fuji: Rename convergence point to stereo window position.

Proposal for Fuji: Provide stereo window position and focus distance in the EXIF data.

35 mm equivalent focal length

The 35mm equivalent focal length of the FinePix REAL 3D W1 is (horizontal field of view):

The 3D 35mm equivalent focal length is larger due to the cropping in this mode. The manual (on page 105) states a zooming in 3D (and ADV 2D) mode of 39 mm to 149 mm, this is thus close to the above 41 mm to 154 mm.
In ADV.3D it is the same as in 2D, so in this mode no cropping.

Proposal for Fuji: Why is the correct focal length not in the EXIF data?

Using the term parallax

Fuji uses the term parallax for the shift to move around the stereo window. IMHO this is a wrong usage of the term parallax (which normally is the difference in viewing angle in the two pictures, due to different distances of objects in the scene).

Proposal for Fuji: How is the auto (parallax) shift determined in the camera?

Proposal for Fuji: Don't use the term parallax for the shift in stereo window, use the term shift!

EXIF data

How would such a crop, shift and resize be made visible in the EXIF data? If the actual effective CCD size used in the picture was available in the EXIF data, one could determine this. Are this PixelXDimension [A002] and PixelYDimension [A003], EXIF 2.2 (Table 4 or page 26) (aka ExifImageWidth and ExifImageHeight) or ImageWidth [0100] and ImageLength [0101]: EXIF 2.2 (Table 3 or page 17)?
I think not; looking at the description in EXIF 2.2 page 17 and 26. I think some kind of transformation matrix is needed (that includes zoom/rotations/shifts/etc.

Other ideas, me know (I have send e-mails to ExifTool, Jeita and Meta Data Working Group).

The CIPA DC-006-2008 standard might also be handy to cover some of these questions.

Proposal for Fuji: Add effective CCD size in pixels for the particular picture (use FocalPlaneXDimension and FocalPlaneYDimension?).

Haziness on a left or right picture

It looks sometimes there is glare/haze on one of the pictures (there was no Sun shine).
Glare on left picture

Might this be due to bright sky glare (again no Sun at this moment) on the plastic that covers a lens? I don't think it are dirty fingers on the lens or dust on inside (as this has been tested with camera directly after an RMA clean-up by FujiFilm),.
Partially it looks to be the standings sides of this glossy metallic plate on the plastic lens cover and/or the shinny on/off switch of the camera. Baffling these shiny sides works very well (see this demonstration) or by using a lens shade I made or a nice surround (pers. comm. M. Pow [2010]).

I am working on a prototype (clip-on) lens shade for the W1; which could reduce:

the dirty fingers on lens
the reflecting light against the standing sides of the glossy metallic plate touching the plastic lens cover
reflecting light from camera's on/off switch
reflecting light from fingers/hand near lenses into lens systems
relfecting light on dust particle in the inside.
reduce UV light
the ghost dots due to the flash and/or
the high contrast glaring/haze.

Proposal for Fuji: Is there a possible solution for this issue (like the above proposed lens shade, if that is the reason)?

Thinness of faces

This happens due to the small focal length of the camera, due to perspective distortion. This is not related to stereo photography, as it is clearly visible on a single 2 picture. Solution is to zoom a little in (two or three stops).

Proposal for Fuji: Mention this phenomena in the manual?

Pincushion and barrel distortion

The barrel correction (pincushion correction if negative) needed in 2D mode is depending on the Focal length (F_c):

Barrel correction depending on focal length

Barrel (when negative; pincushion) strength within PSP,
when compensating pictures made with
Fujifilm FinePix REAL 3D W1 in 2D mode

The distortion looks to be smaller than for FinePix E510 or other cameras.

These distortions are not automatically compensated for in the 3D mpo file. To compensate the barrel/pincushion distortion in the 3D mpo file one first has to resize (depending on the focal length) the mpo file to the original effective CCD size, shift this picture and add borders to get back to the 10 MP (reverse of the cropping) and then compensate the distortion.
A PSP macro script is available to compensate for the resizing, shifting and barrel/pincushion distortion (all under your own risk). Please ask me.

Proposal for Fuji: Compensate the barrel/pincushion before making the MPO file.

As these distortions could cause unnecessary eye strain, an automatic distortion compensation might be handy in the Future of Imaging system; as the FinePix REAL 3D W1 is a monitor (W1), viewer (V1) and part of the print system (or is the compensation done in the V1 and print system?). If people know, let me know.

Optical axis shift

The Optical axis shift parameter (page 87 of the manual) of the 3D W1 determines the vertical position of the left lens.

Proposal for Fuji: Is the lens actually moving or is it just a vertical shift on the file?

On my camera the variability between Optical axis shift and the Position Vertical shift (in SPM) can be seen below. Why this option is a user selectable option I don't understand; it is due to mechanical wear that one needs it in the future?

The vertical shift in the resultant pairs change for my W1 camera (changes per camera: Dmitry Semyonov, pers. comm. [2010]) somewhat when using different focal lengths:

Vertcial shift in REAL 3D W1 when changing focal length

Shooting guide lines for 1/30 rule

Would be handy to have some shooting guide lines in the monitor when making a stereo picture based on ADV. 3D Individual shutter shooting mode, so that one can see if the second picture will be close to this 1/30 rule

Proposal for Fuji: Provide a small grid with a 1/30 division as part of the shooting guide lines.

A hot pixel

Like Michael Brittner reported, my camera also shows on the CCD white-ish hot pixel in 3D mode for both still and video mode (not in 2D mode). This hot pixel is in the right picture of 3D still file or 3D video file (in the video file the single pixel seems to be spread out).
See this 3D video right picture (this was using fully covered lenses, so no light was present):

For F_c=6.3mm the hot pixels are around 1866/725 and for F_c=18.9mm they are around 1682/535 (this change in position is due to the cropping and resizing done in the camera).
The camera has been send for repair to FujiFilm and the hot pixel has been masked out.
After a clean-up activity by FujiFilm, I got a hot pixel on the left CCD:-(

Proposal for Fuji: The quality control should be a little tighter, as hot pixel should be masked out in new cameras.

Experience exchange

Proposal for Fuji: Would be nice if Fuji would facilitate a moderated chat with technical experts in Fuji and the stereo world; to understand and help develop the W1.

Acknowledgments

I would like to thank the following people for their help and constructive feedback: Mike Brittner, Arturo Canalda, Co van Ekeren, John Goodman, Phil Harvey, Jeff Jeesse, Barry Mason, Linda Nygren, Mathew Orman, Andreas Petersik, Bob Pleiff, Max Pow, Brian Reynolds, Dmitry Semyonov, Jan Visser, Gert-Jan Wolkers and all other unmentioned people. Any remaining errors in methodology or results are my responsibility of course!!! If you want to provide constructive feedback, let me know.

Disclaimer and Copyright

Home

Mail

Major content related changes: Oct. 30, 2009