Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

Oops (Patreon)

Published:
2024-06-26 02:18:06
Imported:
2026-06

Content

Thanks to an attentive viewer, I have been informed that I got the entire explanation of Bayer color filters wrong in the gray scanner video. This does a tremendous amount of damage to the narrative and I will probably have to reshoot a big chunk of it. Just wanted to let you know, in case anybody else notices it and wants to point it out; there will be a new version of the video sometime in the next couple days.Thanks go out to that viewer and everyone else who has corrected me over the years, I would look a lot dumber without your help.

Comments

Chas Becht

Happens to the best of us.

John Simon

You might also want to provide conversions to 2024 prices because $6000 PC in 2024 is the cost of some Linus Tech Tips abomination. $6000 in 1991 is new car money

Aaron the Tinkerer

Awesome to hear it was caught, can't wait to watch the update :) Thanks for all your hard work!

Grey Panther

No worries, it means that I get to enjoy the video twice! Once now and once you re-upload it! :)

Scott Kemp

This is the kind of humility I like to see in people. Thank you for the work you do.

Funkmon

I'm a photographer and I think you got the Bayer filter thing right in the vibes. The Bayer filter does let only one color of light through and must be demosaiced. Black and white sensors are therefore pixel level sharp and more sensitive to light, and so are Foveon, at least as far as I know. For your explanation to be so wrong so as to require a substantial rewrite is shocking to me. I am interested to see the difference.

cathoderaydude

The error is subtle until you realize what it implies. The whole bit about needing four times as many photosites to make a color sensor? Completely wrong. I was under the impression that Bayer sensors all do pixel-binning; that each pixel in the finished image is the product of four photosites, and thus it contains luma averaged across those four points, then color interpolated based on a complex matrix. The reality is that if your camera outputs a 4000x3000 jpeg, the sensor only has that many photosites. There is no binning going on, just a ton of interpolation. If you look at the pixel at (5,5) in the un-debayered raw, you'll see a red, a green, or a blue channel that was captured at (5,5) on the sensor, and that's it. That color channel is interpolated into the pixels around it, and the other two colors are derived from pixels around it.

cathoderaydude

Ah, I love patreon's very good UI where I can never remember which key posts the comment. Continuing: My ultimate explanation of _why this causes image degradation_ is not wrong per se. It is true that each pixel is the product of data from multiple physical points on the sensor; I was able to rescue that part of the explanation, fortunately. But the earlier conjecture, about sensors needing to quadruple their luma bandwidth in order to be colorized? Simply wrong, it never happened. The difference between a color CCD and a monochrome CCD (like you'd find in a security camera, say) is just whether there's a color filter array. It's otherwise the exact same sensor with the same number of photosites. The end result of this is that I'm a lot more irritated at the existence of bayer filters than I was before. 3-chip sensors are even more "the right way to do it" than I'd realized - they're so straightforward. Every pixel is the product of... measurement of that point in the image. Plain and simple, no complex math, no interpolation, no guesswork.

Funkmon

It was so subtle I missed that you explained it as pixel binning. I know we're big on technically correct being the best kind of correct so good job. It's just the other side of the coin though, isn't it? I interpreted it as that. The resolution is what it is, but the color resolution is halved due to interpolation in the demosaicing. Your 12 megapixel camera has 12 million photosites, but since each one of those is only one color, it's effectively a lower resolution, presumably about a quarter, interpolated out, so if you wanted a true 12 megapixel perfect image you would really need 36 megapixels in true RGB (Like Foveon). I think Bayer X4 or X2 or some shit Sony put into phones might work like you described though to effectively get pixel level sharpness and lower noise by simply putting in a yes, 48 megapixel sensor.

cathoderaydude

Yeah I mean that's my take - they're still basically binning, just... much worse than I realized. They have really good algorithms for it, the result is really impressive, but it's still binning conceptually. It twists my brain into knots; I can't decide whether it's really reduced chroma or luma or both. I'm pretty sure it's both. Anyway, the end result is that all I really had to do to fix the video was edit out the segment where I assert that they had to increase the resolution of sensors to achieve color. The rest of it's still valid within a margin of interpretation, which is good enough for me, and saved me from doing a rewrite and reshoot which I _really, really, really didn't want to do_. So I'll call it a W.

Aaron Ladd

Another rabbit hole using the same style of sensors are Line Scan cameras used in racing sports timing. they do use a lens, but require special software to use. And the cameras cost more than a new car. I have experience with a brand called finishlynx