now i really want a diffraction grating. hallucinated ssf for the panasonic s9 are sufficiently different between the gaussian model in vkdt mkssf and the autoencoder mentioned above (made it work after deleting some version-update-bitrot in the python scripts).
have to experiment a bit more, but i really like the auto encoder-generated lut so far. seems to match the camera matrix a bit better out of the box. maybe it’s because they disregard all the rest that’s in the dcp (HueSatMap), but maybe that’s for the better? also there’s quite a few dcp for this camera, not sure which one best represents the actual look-free transform.
I used the Edmund Scientific 12mm transmissive grating, 1200 lines per mm, currently $115US. Main challenge with these is holding them, I ended up making a spring-loaded clamp made of stripwood using a ball point pen spring. I also bought the 12mm diffuser, about $15US, rather than messing with wax paper and such.
I also tried the cheap printed gratings, about $19 per box of 10, plastic mounted in 35mm slide holders. They didn’t transmit as much light, made camera alignment challenging, but could yield similar results if you got that right.
You’ll need to collimate the light shiing on the grating, easiest-to-make contraption is two razor blades forming the slit.
In audio circles, people send their headphones to oratory1990, an acoustic engineer from Austria. He will measure them on professional rigs and provide the measurements + EQ (Harman) for the community. But I guess cameras have more inherent variability than high end headphones?
No, the Bayer dyes are spec-ed specifically for a given model, and IMHO for a brand, for some brands. Take a look at the two Canons and the two Nikons on this page:
For the light source, pretty much any white LED can be used. They don’t say anything about the size of the slit in front of the light, so I assume it doesn’t matter too much, as long as the edges are sharp. The slit is 2.5mm
The Amazon one is 500 lines/mm, the ebay one 1000, which affects the transmissibility (500 would be brighter), but I think the higher lines provides greater gradation of the spectrum, ergo, more precise. A licensed optical engineer should clarify/correct.
In terms of handling, the 35mm slide ones are easiest to deal with.
The lines/mm also affects the angle to which the spectrum is displayed offset to the perpendicular centerline, need to know that angle so the spectrum is splayed symmetrically on the camera sensor. 1000 lines/mm requires a 34-degree angle, 1200 lines/mm requires 42 degrees, these corresponded to the cheap 35mm plastic grating and the 12mm ground glass grating, respectively. For other lines/mm, refer to this post for the equation and method:
i spoke too fast about the autoencoder it seems. they don’t consider the HueSatMap in the dcp and maybe regularise a tad much. in the one scene i tested (green tinted daylight under a tree) it worked really well, but didn’t hold up under tungsten illumination. this one here seems to be smooth enough and generally looks very close to the colour matrix processing in easy cases.
maybe a cheap diffraction grating film can give the rough spectral shape and then an optimiser can push it into place instead of calibration of pixel vs wavelength position?
As bad as it may look, I feel incredibly happy for -
being able to actually see how smartphone OEMs just guesstimate their way through this DNA for the image, if I may call it that, making it incredibly hard to get to a good point for many users as well as trying to match it to other sensors.
Being able to tinker my way and understand how sensors work and how the light it sees is converted all the way into a pretty looking image for my eyes.
being able to achieve a somewhat usable profile for both in camera and post production, although with the help of LLMs, that actually is reliable and less dependent on the fancy tonecurves of Adobe etc (they do have DCP profiles for Pixels it seems).
Maybe I should move to your p4 of your series next once I redo the setup with as much care as I can put into it starting from maybe using blades to make the slit as sharp as I can (the cardboard I cut is far from perfect). Then making the CD piece a lot more stable but flexible enough to be calibrated. Also other parts of your own project that I skipped for this test.
I also played around with the halogen inside the bike’s headlight but it seems to be a bad idea because they may have some coatings or something else going. I will use a Tungsten bulb next along with my CFL. Avoiding Noise as well as any shifts in the spectrum are priority because in this test I had a massive 20nm shift with one end. Couple that with the other fragile parts and you get what I shared.
Makabe approach yielded me instant results with just a white CFL bulb and it wasn’t at all bad but it lacked Blue channel data for some reason yielding pink skins and bricks. Was pretty good for making me feel more confident into investing more time into the approach now I official call “ggbutcher” approach.
Maybe most of this is just rookie words but I am very happy and excited!
Max dE of ~6 on the red patch, really, kinda in line with matrix performance, although I think you’d want a better dE for one of the primary colors.
Yogansh, this is really cool results. My Calibrite spectrometer is in transit; one of my first measurements will be my CFL bulb, with the intention of resurrecting data from pt 4 and attempting to make a profile with just the calibration image. I’ll be posting the whole attempt here, probably as a pt 5…
Oh 100% that’s the case. I think the setup has only way up from here so I should be able to squeeze a lot out of this if nothing goes wrong.
I will be anticipating your post!
I initially wanted to do something similar to this. (Thanks to Tadi )
This might turn out to be a very convenient way to build a database for countless bricks designed by people who don’t care about them beyond net margins.
Y’know, the hardest part of all this appears to be getting cameras in front of spectrometers/monochromators. Reminds me of the Kennedy Space Center bears, a male who lived on the north end of the reserve, and a female who lived on the south end. Never got together, kinda sad story…
I think the OEMs are holding both of them in their hands and still not realizing that the two pieces they are holding are part of the same puzzle.
I love the analogy. there was a lack of a human who cared enough or maybe they all assumed nature will take it’s course. We are left at nature’s mercy in this tech jungle I suppose.
Mother nature gave us the power of reasoning so that we could make better decisions but I guess that only applies to survival yes no decisions : )
I’m curious why are fujifilm stocks producing a fairly strong blue cast when you preflash through them? Other stocks seem to be reacting correctly with no casts, but with the fuji stocks the preflash is practically useless
)