I have additionally verified that some recent fisheye lenses exhibit unusual TCA patterns that may prohibit from totally efficient full TCA correction on the full digital sensor coverage, especially when these lenses are "abused" by (shaving) and/or adaptation on Full-Frame DSLR.
Comparative plots and illustrations of test results are presented that exemplify TCA.
Some suggestions are proposed in the conclusion.
This problem affects the image quality especially in the outer part of the image as TCA increases with the radial distance from the image center (that is the only point where it is nil).
This is worsened by the use of fish-eye lenses, when the angular coverage becomes very large (typically 120 to 200 degrees). TCA then swells up toward the corners of the captured image (or the edge of the circular image) and hampers the quality as well as lessens the sharpness of the image in these critical area.
TCA is tamed by the designer of the fisheye lens so as to be tolerable in the full coverage on the camera sensor. The most frequent sensor size is APS-C: outside of this capture area, TCA could be quite intense there but without any practical consequence as the light falls outside of the useful sensor capture zone.
Some fisheye lenses where originally designed to be used indifferently on FF (film or digital type) or APS-C while other recent model were designed for exclusive use on APS-C camera. However some of us panographers have modified (aka shaved) and mount adapted these latter fisheye lenses to work on Full Frame DSLR (e.g EOS FF DSLR series) even if they were not intended to be. This opens the way to lot of creative applications or allows previously impossible panorama realizations, but consequently TCA then has generally to be corrected on the image before stitching in this case of "lens abuse".
All fisheye lenses suffer from TCA but they are not equal in that matter.
TCA (aka lateral color) results in a shift of the pixels of the red and blue channels mapping with respect to the green channel that we generally for different reasons (but arbitrarily) consider as the reference channel.
It is sometime rather easy to -almost- completely correct the aberration by means of simple linear correction of the red and blue image channel shift. One of the commonly used tool is Adobe Camera Raw (ACR) generally associated with Photoshop. The correction can be conveniently done upstream in the workflow i.e. together with raw image conversion.
Many have saved the conversion profile that is specific to (Camera + Lens) and subsequently proceed to batch processing of the source images prior to inputing in a stitching software. I shall demonstrate in this article that this can be done only with adequate care and caution.
Additionally, I have incidentally verified that TCA for some lenses cannot be fully corrected with even the most sophisticated tools under some specific circumstances. As a consequence, some "deep blue fringing" may resists all attempt to get rid of them. This is not caused by photosite overflowing as it may commonly be mistaken while it can be obviously be mixed with this other problem along some high contrast object edges. Indirectly and additionally affecting sharpness, this is however mainly visible only at high zoom-in level image examination.
More details on this subject have been presented in this supplement "Image from a Nikkor 10.5mm: a case study for C.A. analysis"
Claiming that the nature of the light on a photographic scene may have an influence on the TCA is a plausible conjecture as one can assume that different types of illumination may impact this amount of pixel radial shift differently. In short: TCA is obviously IMHO a wave length dependent lens error.
In the course -but as afterthought- of an ongoing more general comparison testing testing of different fisheye lenses, I have also paid special attention to this point as I could not find much prior information related to fisheye lenses by extensive googling...
Furthermore I have succeeded more or less in testing some fisheye lenses under natural sun light (in fact it was an overcast ideal sky that day, no shadow at all, and this helps a lot).
Hereafter is a preliminary report of my findings that I believe corroborate the previously stated conjecture.
The main feature of this test set-up is a rather shallow cylinder that is lit by an electric light bulb located on the longitudinal axis of the closed cylinder. This can be either tungsten or fluorescent lighting for the moment. I hope that multi-LED shall complement the limited present alternative in a near future, as this very new type is supposed to be the closest possible simulation of natural sun light that is currently accessible to my limited buying power.
A specially designed target is stuck on the inner wall of the cylindrical support. In addition to a classical lens testing patterned target, two types of "checker ribbons" were inserted: one (top) is made of matte black and white paper rectangles while the other (center) was printed on a glossy EPSON roll of paper.
The fisheye lens under test is also located on this axis right under the light source: the entrance pupil for 90° incidence angle is accurately placed there as the lens is mounted in front of my Canon EOS 5D. Direct illumination of the lens front by the light bulb is blocked by a circular shield that otherwise evenly lets illuminating the cylindrical target at an angle that avoids any direct stray light reflection from the target toward the lens itself.
This circular test bench is isolated from external stray light contamination by means of a black fabric cylindrical booth. Some of the flat surfaces of the apparatus are additionally covered with light absorbing deep black seersucker paper.
QTVR movie of the test set-up.
Lenses submitted to test:
1) The newer Sigma 8-mm f3.5 that I too briefly had in my hand for test was unfortunately not subjected to tungsten lighting test evaluation nor was it tested under natural sun light:-(
2) The Canon and Zeiss lenses were not subjected to test in sun light illumination.
3) The Tokina has been measured for every millimeter of focal lens from 10 to 17 mm, but only two typical values shall be presented here (10-mm and 16-mm)
The optical main lens axis is adjusted and aligned to aim at the ZERO -center- mark of the target and at an altitude that makes the checker ribbon "flat and straight" on the image: the ribbon is in the "equatorial plane".
The .CR2 (Raw images) files were processed by applying scripts developed by Pablo d'Angelo (Thanx again Pablo). BTW: He had already co-authored a closely related article.
Dcraw shall be mandatorily be invoked downstream of the workflow. Incidentally, this S/W produces an image that is not deprived of any bordering image element (dcraw yields image from all the full EOS 5D active elements of the sensor i.e. 4386 x 2920 pixels to be compared to the standard -cropped but official- 4368 x 2912 pixels count). Consequently, dcraw is manually used first to view the test image in order to define the coordinates of the "experiment line" that is drawn across the entire checker ribbon from one corner of the image to the diagonally opposite corner. Photoshop is used for this manual operation.
From a script hand-written under XCode (MacOS X) that is invoked by another Octave master script through Terminal (MacOS X), a resulting output linear 16 bits image (dcraw produced) is then subequently interpreted by a set of tools that are Octave S/W driven. Most are interpreting the transition (from black to white checkers) properties and corresponding illumination R, G, B levels along the "experiment line".
Graphs are subsequently prepared under GNUPlot and a few following examples are illustrated hereafter.
Beware: The Y axis scaling is very different from one plot to another!
Animation: You may fly over each of the plots with the mouse to view a simulation of the approximative coverage limits on APS-c sensor (portrait mode)
Then click and hold down to view alternatively these limits for landscape mode.
--> I am sorry that I could not make these animation to play smoothly especially with MS Internet Explorer/Vista: The landscape mode takes a ...long time to load :-(
|Sigma 8 mm f4.0
|Sigma 8 mm f3.5 (new)
|Tokina 10-17 mm -> 10 mm @f/5.6|
|Nikkor 10.5 mm f2.8
|Canon 15 mm f2.8
|Tokina 10-17 mm -> 16 mm
|Zeiss 16 mm f2.8 @f/5.6|
1) The reader may compare the graphs to assess the relative propensity of each lens to produce TCA affected images at f/5.6. I shall present another more exhaustive comparison ASAP.
2) The first part of the following conclusions are applicable to Full-Frame DSLR (i.e. only some Canon EOS D cameras at the time of writting).
RED channel: The impact of the change on the red channel is not much when switching the light from Tungsten to/from Fluorescent. In addition, experience has taught me that image red channel TCA is easily correctable for all the tested lenses.
BLUE channel: On the contrary there is a very obvious large difference on the blue channel when exchanging illumination types. The fluorescent light exacerbates dramatically the radial shift of the tungsten blue channel image pixels. This is applicable for all lenses. The extent of this effect has been a personal astonishing discovery as more that one to two pixels difference can be observed near the edge of FF coverage.
Natural sun light: On both red and blue channel, the illumination by natural sun light yields almost an identical result as the fluorescent lighting, the latter being slightly worse. (See Sigma and Nikkor plots above).
It is therefore not reasonably possible to use the same settings for TCA correction when the lighting of a scene is changed radically from one type to another especially when tungsten is involved. From my experiment, this may however be possible when going from outside photography to artificial (economical type) fluorescent bulb lighting. This has however to be confirmed by further experiment.
Another observation can be made: the shape of the curves describing TCA for the Nikkor, the Tokina, the Zeiss and for the Sigma 8-mm f3.5 (differently and in a much less extent that is barely perceptible here above for the latter) exhibit a reversing gradient on the blue channel at around 60 to 70° (the plot thus resembles the typical appearance of the extended wings of a bird of prey which tip feathers bend upwards). This is strongly amplified with fluorescent lighting and this IMHO yields the impossibility to fully correct TCA with currently available high grade tools (e.g. PTShift). The older Sigma 8-mm f4.0 and the Canon 15 mm f2.8 behave differently as TCA is concerned and is near-fully TCA correctable even with standard linear tools (e.g. Photoshop ACR). As I personally do not own neither the Sigma 8-mm f3.5, nor the Canon or the Distagon Zeiss lenses, I cannot make sure that this property also applies to these model, but it should according to my understanding...
APS-c DSLR case:
As the APS-c sensor is about half in area and as the linear dimensions are 1.5 to 1.6 times less than FF, the impact of the peculiarities presented above in my conclusion should be amended.
Most of the TCA "defective" part of the image (e.g. the strange blue channel image behavior) is cropped from the captured APS-c image. Only small areas (near the corners of the source rectangular image) shall be visually TCA affected and therefore possibly additionally impacted by the nature of the light wavelength. TCA is then much less a concern when using the fisheye lens on APS-c Digital cameras.
The "blue weird" problem may still subsist though and for example bare tree branches against a very intensely white sky (over-exposed or not) and at about 60 degrees or more of incidence angle from the lens longitudinal axis may be "border-lined" by this blue artifact
BTW, the two "shaved" lenses that show the blue channel "problem" with the more obvious manner, were explicitly not intended for FF DSLR use by their designers!
Finally, I should summarize that in any case and after proper TCA correction, even with FF camera, only a very small part of the remaining TCA "defective" images is visible in the output panorama when a sufficient number of overlapping source images is taken into account for stitching: the TCA "defective" remaining portions should be cropped and overlaid by image of much better quality during the stitching and blending process.
30 June 2007