Hacker News Reader

Comment by rocqua | Hacker News Reader

> So the novelty is that they separate R,G and B planes before compression?

No. The novelty is that they compress before de-mosaicing (aka debayering).

In other words, rather rebuilding the image from the signals and compressing that image, it compresses the signals themselves (not even seperated I think).

View on HN

I'm just guessing it makes more sense to separate them. In my naive no knowledge version. if it was

   RGRGRGRG
   GBGBGBGB
   RGRGRGRG
   GBGBGBGB

Then, if I compress losslessly just grayscale I know at the other end I just interpret the first pixel as Red, next as green, red, green, etc., Then on the next row I interpret the first pixels as green, next as blue, next as green etc...

BUT, if I compress lossily then the unrelated colors next to each other are going to bleed into each other while compressing. VS if I separate them out so

   RGRGRGRG
   GBGBGBGB
   RGRGRGRG
   GBGBGBGB

becomes

   RRRR BBBB  
   RRRR BBBB

   GGGG
   GGGG
   GGGG
   GGGG

The compression would make more sense and less bleeding of the signals from each color into other colors.

It would also seem to compress better. Imagine you're filming something pure green. Exaggerating but you'd get this

Which after separation is

That's just a guess though.

A different comment suggested they compress the green channel, and then compress the red and blue channel as differences against the green channel.

This makes a lot of sense to me, since green is sort of 'luminance'. It is like a poor-mans LAB color space that exists before demosaicing.

This is a very good guess, and I believe it's precisely how it's done. Doing anything diff-ing the color channels still requires a lossy transformation step, so it's not quite as "raw."

And you're right, that treating it as a single monochrome image makes it less compressible, because you get lots of high frequency information between adjacent pixels with different color. Which is sort of the opposite of what you want.

Note also that, despite calling it "raw," Red's implementation is lossily compressed. You can even tune the compression ratio of their "raw" codec, and back in the very early days, there was certain subject matter that just looked terrible coming out of the camera, even as raw -- I remember one particular tree that an early user would film to test new firmware updates; in the earliest firmware builds, it actually crashed the camera (!). The images it did produce showed tons of wavelet compression artifacts.

Reply on news.ycombinator.com