3D-Printed Anaperture Single-Shot Anaglyph Aperture

ProfHankD wrote:

Oleg L K wrote:

Wow! Two "heavy-weights" of stereo-3d posting in one thread !

Regarding the single-shot anaglyph approach, I must admit I didn't try the anaperture. But I did experiment with DIY analogs of QDOS lens - I used to glue red-blue filters on the back of (generic) 35/2.8, Minolta 24-104 f/3.5-4.5 and Minolta 35-70 f/3.5-4.5. All are full-frame lenses used with APS-C cameras.

The filters shaping the OOF PSF need to be positioned such that they form an aperture rather than vignetting, and for most lens designs there are generally several potentially appropriate places for that. I prefer in front of the lens because it is easier to access, but behind the lens is also possible in some cases. Replacing the original iris of the lens always works, but is very inconvenient....

I confirm the bokeh being odd, but generally the results are interesting.

The bokeh are only odd if the two color filters aren't resulting in identical size and shape apertures. For example, just making opposite sides of the aperture different colors produces the "misaligned" bokeh seen with the Q-DOS lens.

In any case the area where both left- and right frames appear with proper "strength" is larger when the lens diameter is larger - e.g. faster aperture and/or larger focal distance.

Actually, it's not about how fast the lens is, but how little it vignettes.

In my opinion the most appealing is 35/2.8 (or faster), since "normal" lens is desirable for stereo, not telephoto. 35/3.5 and 29/2.8 proved to be unusable.

The key problem is that the color-coded aperture needs to be smaller than the largest non-vignetting aperture -- which is usually 2-3 stops down from wide open. With a 35mm f/2.8 lens, that would typically mean around f/5.6, and 35/5.6=6.25, so you'd have to fit both colored apertures within a 6.25mm diameter opening. That can be done, for example using two 2.75mm diameter circular openings separated by 0.75mm (baseline of 3.5mm), but that's pretty touchy for homemade stuff. The resulting effective aperture would be f/12.7... which is actually still ok for FF (or even APS-C, but would be past the diffraction limit for typical MFT).

Correct me if I'm wrong, but the main catch with this approach applies to anaperture too: you cannot make ghost-reduced anaglyph. E.g. if your scene doesn't include saturated primary colors, the result is fine, otherwise the photo is non-viewable.

Fundamentally, all dye-based filters leak. Thus, there is always some ghosting... but there also is potentially enough information to computationally correct it. I got pretty far along on the correction back around 2011-2014, see Reprocessing Anaglyph Images . However, there was significant interest in the technique from some key movie-industry players at the time, and I was never able to produce full-color stereo pairs with color quality sufficient to make cinematographers happy (which is a VERY high standard).

If I was to do this now, I'd probably combine the "color guessing" algorithm I originally used with some AI methods as well as the more typical techniques (e.g., deconvolution and stereo matching, which by themselves produced inferior results). Here's the really short explanation of the color guessing algorithm from that paper:

In a typical anaglyph, some fraction of the scene is naturally aligned to within a pixel. For example, a baseline of a few inches is negligible relative to a background that is miles away. If the naturally aligned regions can be recognized as such, the color information they provide is known accurate. Not only is it accurately describing the full color of each aligned pixel, but it also is establishing that those specific RGB triples actually occur in the luminant-colored scene. For example, if RGB value 0x802a1c is known to occur frequently in the aligned portion of a scene, it seems likely that a pixel in a not-aligned region with R=0x80 and B=0x1c should have G=0x2a.

Suppose we are examining a region of an anaglyph which has a relatively constant color for a horizontal distance which exceeds the maximum possible stereo misalignment between left and right views. The object in this region might not be fully aligned – there may be color fringes on either or both sides of it. However, the colors that appear on the wide overlapping portion of the object are approximately correct. In fact, if there are fringe colors on either side of this region, the actual alignment is most likely a shift corresponding to the width of the fringe, so even more precise color samples can be obtained.

The identification of color fringe areas is easy for humans, and approximately identifying them can be fairly easy for a computer. A simple approach is to look for areas with colors that are heavily biased toward either of the anaglyph coding colors; for example, in a red/cyan anaglyph, strongly red or strongly cyan pixels are good candidates for fringe areas. More accurately, a region with an RGB color tuple in the anaglyph that is relatively far from any RGB tuple known to occur in the image is probably fringe. The accuracy can be further improved by imposing the rule that the pixels can only be fringe if they are in a horizontal group less wide than the maximum shift. We do not know the true colors of fringe areas, but it is likely that they do not introduce color tuples that are dramatically different from any seen in other portions of the anaglyph.

I generally focus my research on things that seem to be most useful to the community, and frankly the interest in anaglyph capture was way lower than I expected... or wanted a level of color accuracy that I couldn't find a way to deliver. Cinema folks are VERY picky about color.

I could easily pick this work up again, but lacking both research funding and public interest for it, that's very hard to justify....

Hank, I would say that sans funding and based solely upon your personal interest you accomplished your goal admirably!,,

Dave

What would really be nice, is a device that takes full-color left-right photos sequentially through anaperture holes. I'd think of shutter-glasses synchronized with 30fps 4K video capture.

Actually, if Canon hadn't botched the dual-pixel stuff so badly, it would be pretty easy to use it to capture a full-color high-quality stereo pair at a time. I bought a 5D IV as soon as it came out just to test using the dual-pixel raws as an aid to this. Unfortunately, the print-through on Canon's dual pixels is giving only about 2 stops of SNR, and they also botch the exposure computation so by default you get pixels clipping. I'm amazed Canon has even gotten PDAF to work so well with such poor dual-pixel data....

Earlier this year, at Electronic Imaging 2021, I presented Programmable Liquid Crystal Apertures and Filters for Photographic Lenses . I had hoped color LCD panels would be useful as programmable filters, but my experiments say nope. However, it would be pretty easy to use a pair of LCLV to do shutter-synchronized left/right apertures without moving parts (e.g., using the flash trigger signal from the camera to detect the shutter firing or having an external controller control both the camera and LCLV). I found about 9 stop usable DR for the LCLV, so you'd still get some ghosting with high-end cameras (e.g., Sony A7 series) that have 13-15 stop DR, but straight encoding of JPEGs only codes about 9-10 stops DR... so non-tone-mapped (e.g., Sony DRO off) images shouldn't show ghosting. BTW, EI2021 hasn't posted the above paper yet, but my Arduino driver for LCLV is at https://github.com/aggregate/LCLV .