Sigma Fp review and interview / Cinema DNG RAW

[PannySVHS]

9 hours ago, paulinventome said:

Having said that by the time the image is debayered into a working colourspace that 8 bit source of thee RGGB channels is tonally more spread meaning that effectively the end result will have more tonality than 8 bit implies, say compared to an 8 bit movie. I hope that makes sense?

Paul

 

RGGB, sounds interesting, so it's 256 square 4, around  4 billion colours instead of 16 million colors? Anyway, keep us posted please. Sounds like an awesome read to come.

[KnightsFan]

10 hours ago, paulinventome said:

Having said that by the time the image is debayered into a working colourspace that 8 bit source of thee RGGB channels is tonally more spread meaning that effectively the end result will have more tonality than 8 bit implies, say compared to an 8 bit movie. I hope that makes sense?

Doesn't make sense to me. 8 bit bayer is significantly less information per pixel than 8 bit 4:2:0. RGGB refers to the pattern of pixels. It's only one 8 bit value per pixel, not 4.

[PannySVHS]

@KnightsFan fun thing: you had 444 likes. i changed that.  

[KnightsFan]

6 minutes ago, PannySVHS said:

@KnightsFan fun thing: you had 444 likes. i changed that.  

Haha thanks, I didn't even notice that! I have now passed full subsampling on EOSHD.

[Lars Steenhoff]

2 hours ago, KnightsFan said:

Doesn't make sense to me. 8 bit bayer is significantly less information per pixel than 8 bit 4:2:0. RGGB refers to the pattern of pixels. It's only one 8 bit value per pixel, not 4.

8bit bayer is debayered to 8bit 4:4.4

what Im I missing?

[KnightsFan]

39 minutes ago, Lars Steenhoff said:

8bit bayer is debayered to 8bit 4:4.4

what Im I missing?

It can be, yes. I was responding to the claim that the fp's 8 bit raw could somehow be debayered into something with more tonality than an "8 bit movie." I suspected that @paulinventome thought that RGGB meant that each pixel has 4x 8 bit samples, whereas of course raw has 1x 8 bit sample per pixel.

My statement about 4:2:0 was to point out that even 4:2:0 uses 12 bits/pixel, thus having the potential for more tonality than an 8 bit/pixel raw file has. Of course, 8 bit 4:4:4 with 24 bits/pixel would have an even greater potential.

[kye]

https://noamkroll.com/why-i-just-bought-the-new-sigma-fp-mirrorless-camera-for-filmmaking/

[paulinventome]

On 10/20/2019 at 12:57 AM, KnightsFan said:

It can be, yes. I was responding to the claim that the fp's 8 bit raw could somehow be debayered into something with more tonality than an "8 bit movie." I suspected that @paulinventome thought that RGGB meant that each pixel has 4x 8 bit samples, whereas of course raw has 1x 8 bit sample per pixel.

My statement about 4:2:0 was to point out that even 4:2:0 uses 12 bits/pixel, thus having the potential for more tonality than an 8 bit/pixel raw file has. Of course, 8 bit 4:4:4 with 24 bits/pixel would have an even greater potential.

Sorry, no what i mean is that when the debayer happens then missing pixels are reconstructed, usually in a theoretical colourspace like XYZ, then there's a matrix which brings that back into a white balanced space. All that math happens in floating point. So as we say in each 4x4 block there are two G and RB. Some sensors use different green filters as well, so one of those greens might be more sensitive to light than the other. The point of the debayer is to make up for each pixel what the other two missing colour values would be. This is not a simple extrapolation but can get quite complicated. So when that 8bit RAW source is debayered then it will be debayered into a higher bit depth container, or in most grading apps, a floating point linear colourspace. So those 'basic' 8 bit values once gone through the process would end up tonally in a different place entirely, especially the reconstructed channels. This is the point of bayering - we're making stuff up that's not there. Does that make sense? It's a very very different beast to working with an 8 bit 422 baked movie source. If you grab the 8bit UHD sample and push it around you'll see that it's way more robust than it should be and in fact if you look on a waveform (assuming workflow is correct) then you'll tell that the resulting image is not 8 bit.

As for the movie recording. This is the IMX410 sony sensor. This can read 6K at 30fps, so perhaps sigma are reading the whole sensor and making up a RAW from that (bit odd but can be done). Or that sensor does support 4K crop at higher bit depths and rates. I do wonder whether sigma when set to full frame does the whole sensor but if you set to crop then sets the sensor in that mode and can achieve better rates

Finally what we don't know is that whether the 10 bit version of DNG is linear or not. If they've done the same thing as they did with 8 bit then the 10 bit DNGs would be all you need. No need for that extreme 12 bit mode because i doubt you'd see any difference. So fingers crossed sigma have not done linear 10 bit. AFAIK the black magic DNGs are all 10 bit log internally. I don't know if that's still the case as i'm not BMD based (have a Red)

cheers
Paul

 

[paulinventome]

On 10/20/2019 at 12:57 AM, KnightsFan said:

My statement about 4:2:0 was to point out that even 4:2:0 uses 12 bits/pixel, thus having the potential for more tonality than an 8 bit/pixel raw file has. Of course, 8 bit 4:4:4 with 24 bits/pixel would have an even greater potential.

Just to follow up on this. Most of these formats encode in YUV which is where the sampling comes from. An 8 bit YUV file is awful for tonality. The colour difference channels are so lacking detail and it's really just the luma part that 'holds' the image. But, a bit like raw, we don't see YUV or RAW files, these are both decoded into RGB space which we see on the monitors. Whenever there is an operation to 'decode' which takes place in a higher depth, then the resulting bit depth can be higher. If you have any applications that let you see inside a YUV H264 file then you really should take a good explore. When you see the data that is used to store images - be it YUV or RAW then you can see where the quality differences come from. The nice thing about DNGs is no compression, makes a wonderful difference to the image!

You can get a 12 bit YUV file but most are 8 bit. That's 8bit for luma and 8bit for each of the colour difference channels but because of the nature of how colour difference works most of those 8 bit containers are empty. The colour difference channels are subsampled, so if the luma is 1920x1080 then in 422 each of those colour differences are 960x540. But you can decode an 8 bit YUV into 12 bit space if you want...

cheers
Paul

[KnightsFan]

4 hours ago, paulinventome said:

You can get a 12 bit YUV file but most are 8 bit. That's 8bit for luma and 8bit for each of the colour difference channels but because of the nature of how colour difference works most of those 8 bit containers are empty.

I wasn't talking about 12 bit files. In 8 bit 4:2:0, for each set of 4 pixels in 8 bit 4:2:0, you have 4x 8 bit Y, and 1x 8 bit U and 1x 8 bit V. That averages out to 12 bits/pixel, a simplification to show the variation per pixel compared to RGGB raw.

It seems your argument is essentially that debayering removes information, therefore the Raw file has more tonality. That's true--in a sense. But in that sense, even a 1 bit Raw file has information than the 8 bit debayered version won't have, but I wouldn't say it has "more tonality."

4 hours ago, paulinventome said:

Whenever there is an operation to 'decode' which takes place in a higher depth, then the resulting bit depth can be higher.

I don't believe this is true. You always want to operate in the highest precision possible, because in minimizes loss during the operation, but you never get out more information than you put in.

 

It's also possible we're arguing different points. In essence, what I am saying is that lossless 8 bit 4:2:0 debayered from 12 bit Raw in camera has the potential* to be a superior format to 8 bit Raw from that same camera.

*and the reason I say potential is that the processing has to be reasonably close to what we'd do in post, no stupidly strong sharpening etc.

 

 

About this specific example form the fp...

4 hours ago, paulinventome said:

If you grab the 8bit UHD sample and push it around you'll see that it's way more robust than it should be and in fact if you look on a waveform (assuming workflow is correct) then you'll tell that the resulting image is not 8 bit.

I didn't have that impression, to be honest. It seems close to any 8 bit uncompressed image.

 

[Brian Williams]

2 hours ago, KnightsFan said:

but you never get out more information than you put in.

Even if we are including the interpolated data in this statement? If you throw 8-bits into a 12-bit container, couldn't the results (including the newly created data) be an improvement over the original if the interpolation is really good?

[slonick81]

7 hours ago, paulinventome said:

The colour difference channels are so lacking detail and it's really just the luma part that 'holds' the image.

Color saturation is defined not by variety of values of any single (R, G, B) channel arcross the image but by difference between channels in given pixel. Thus color fidelity is defined by amount of steps this difference is digitised. If you have image with (200, 195, 205) RGB values in one point, (25, 20, 30) in other, (75, 80, 70) in third that doesn't mean the image color range is 25-200, it's 10 and it's lacking color badly.

[KnightsFan]

42 minutes ago, Brian Williams said:

Even if we are including the interpolated data in this statement? If you throw 8-bits into a 12-bit container, couldn't the results (including the newly created data) be an improvement over the original if the interpolation is really good?

If we are talking fidelity, not subjective quality, then no, you can't get information about the original scene that was never recorded in the first place.

I'm not sure what you mean by interpolation in this case. Simply going 8 bit to 12 bit would just remap values. I assume you mean using neighboring pixels to guess the degree to which the value was originally rounded/truncated to an 8 bit value?

It is possible to make an image look better with content aware algorithms, AI reconstruction, CGI etc. But these are all essentially informed guesses; the information added is not directly based on photons in the real world. It would be like if you have a book with a few words missing. You can make very educated guess about what those words are, and most likely improve the readability, but the information you added isn't strictly from the original author.

 

Edit: And I guess just to bring it back to what my post was about, that wouldn't give an advantage to either format, Raw or YUV.

[Brian Williams]

Modified Sigma's smaller grip to allow it to hold a T5; haven't printed it yet so not so sure how comfortable it is.

[deezid]

Quite impressed about the RAW samples also. Interpreted them as BMD Film and used my Soft film LUT to convert to Rec709, but made it even darker with way more green - also added some grain und gaussian blur. Really liking the result. 

 

 

[leslie]

8 hours ago, Brian Williams said:

Modified Sigma's smaller grip to allow it to hold a T5; haven't printed it yet so not so sure how comfortable it is.

nice idea mounting it to the side. i presume there are no ports on that side ?

[paulinventome]

14 hours ago, KnightsFan said:

If we are talking fidelity, not subjective quality, then no, you can't get information about the original scene that was never recorded in the first place.

That is the very point of bayer sensors, no? I don't know whether you're a developer or have had a chance to experiment with different debayer techniques but that is precisely what is happening for the 2 pixels for every pixel that are not recorded. The reconstruction algorithms are really clever, taking into account edges and gradients.

A simplistic example:

Red for pixel A is code value 11 and pixel B is missing and C has 14. The debayer can look at the surrounding red pixels and also the colour for the pixel it does have and determine that using a gradient the missing pixel is 12.5. Now you have code values that are not 8 bit anymore.

In addition, these source values are not in a colourspace that we see, they have to go through a matrixing process which also includes white balancing AND highlight reconstruction. These are other steps that fill in more detail.

You cannot do decent highlight reconstruction from a baked 709 YUV image, believe me, i've had to try. You need pre white balancing raw data to do that properly.

You could argue that you're scientifically right but we have been using reconstructed images for decades and YUV is also a reconstruction of the original scene too.

18 hours ago, KnightsFan said:

I wasn't talking about 12 bit files. In 8 bit 4:2:0, for each set of 4 pixels in 8 bit 4:2:0, you have 4x 8 bit Y, and 1x 8 bit U and 1x 8 bit V. That averages out to 12 bits/pixel, a simplification to show the variation per pixel compared to RGGB raw.

I assumed you were talking 12bpp and i see where you're coming from. But from another perspective the U and V is described using an 8 bit range, even though they are packed differently.

At the end of the day the proof are in the images. That 8 bit RAW from sigma is much better than it has any right to be and i spend lots of my days deep in pixels - from Red to Arri and the other end of the spectrum as well. I hate YUV with a passion because it's caused me so many post headaches. I love RAW for its simplicity and flexibility - there are techniques for working on the bayered data before reconstruction or the ability to white balance after the effect in grading means that you can use white balance as a secondary and protect skin whilst changing colours around - it can be way more effective than pulling a key in a restricted colourspace.

What i really hope is that sigma have done the same thing with the 10 bit files, and not made them linear - if that's the case then we're golden.

I assume this is the IMX 410? So it may be doing 6K full frame internally. If you switch to crop mode i wonder if it changes the sensor to 4K crop too - in which case you can pull more depth off the sensor in that mode?

cheers
Paul

[Lars Steenhoff]

Its nice to be able to change the colors to the extreme

[kye]

3 hours ago, Lars Steenhoff said:

Its nice to be able to change the colors to the extreme

Yes, that holds up very well..  even with very subtle tones on her face there's no visible banding or artefacts (even on your upload).  

[Brian Williams]

7 hours ago, leslie said:

nice idea mounting it to the side. i presume there are no ports on that side ?

Don't believe so, and its the same hand grip (HG-11) that Sigma sells for the camera, just fatter now.