The "small raw" output of the 1D III

The Canon EOS iD mark III offers an image output called "small raw" (sRaw).

Do we have any idea what this is? It is said to have about "1/4 the pixels" of a regular raw output, but of course a regular raw output doesn't have any pixel outputs at all - just photosite outputs (associated with places where we will put pixels after we guess what they are by demosaicing). I assume that what they mean is that after development, that output will yield a 1936 px x 1288 px image (vs. 3888 px x 2592 px for the regular raw output). This is the same pixel size as the "small" JPEG output.

It is interesting that the estimated file size for the sRaw output as stated by Canon in the manual is about 7.6 MB vs. 13.0 MB for the "regular" raw output.

Doug_Kerr said:

The Canon EOS iD mark III offers an image output called "small raw" (sRaw).

Do we have any idea what this is? It is said to have about "1/4 the pixels" of a regular raw output, but of course a regular raw output doesn't have any pixel outputs at all - just photosite outputs (associated with places where we will put pixels after we guess what they are by demosaicing). I assume that what they mean is that after development, that output will yield a 1936 px x 1288 px image (vs. 3888 px x 2592 px for the regular raw output). This is the same pixel size as the "small" JPEG output.

Click to expand...

I guess the options are (for each of the R, G, and B sensels):

• only 1/4 of the sensels are sampled,
• the output of 4 sensels are averaged (after applying standard amplification), or
• the output of 4 sensels are added (after applying 1/4 of the amplification), or
• ... ?

Doug_Kerr said:

It is interesting that the estimated file size for the sRaw output as stated by Canon in the manual is about 7.6 MB vs. 13.0 MB for the "regular" raw output.

Click to expand...

My guess is that there is a certain amount of overhead in the file for EXIF info and the preview jpeg, and this won't differ in RAW or sRAW. Lets say that this accounts for 1 MB of the file (I have no idea of whether this is even in the ballpark of the actual value), leaving 6.6 MB vs 12 MB.

Each sRAW value represents a larger part of the image than each RAW value, thus leading to greater point-to-point differences in the sRAW image. In turn, this will lead to poorer compression of the image data. Thus, the remaining difference in file size could be attributable simply to the compressability of the image.

Hmm, would be interesting to know more about what they're doing.

E.

Hi, Elliot,

Elliot Silver said:

I guess the options are (for each of the R, G, and B sensels):

Click to expand...

Ooh! Sensels. I love it!

My guess is that there is a certain amount of overhead in the file for EXIF info and the preview jpeg, and this won't differ in RAW or sRAW. Lets say that this accounts for 1 MB of the file (I have no idea of whether this is even in the ballpark of the actual value), leaving 6.6 MB vs 12 MB.

Each sRAW value represents a larger part of the image than each RAW value, thus leading to greater point-to-point differences in the sRAW image. In turn, this will lead to poorer compression of the image data. Thus, the remaining difference in file size could be attributable simply to the compressability of the image.

Click to expand...

Yes, I could indeed be something like that. I had forgotten about the preview(s). But I assume they would be smaller here, too (1/4 the pixels).

Now, how about this: we combine the outputs of four sensels to make a "fat sensel", and represent that not in 14 bits but rather in 15 bits. (Remember, if we add four values that are known to 14 significant bits each, the precision of the sum is (I think) 1 bit greater. I have to work on that a little!)

Then, for 1/4 the number of sensels, we have more than 1/4 the amount of uncompressed raw data. (But still not enough to come close to explaining the 7.6 MB file size).

Hmm, would be interesting to know more about what they're doing.

Click to expand...

Indeed!

Thanks for your thoughts.

Best regards,

Doug

Doug_Kerr said:

Now, how about this: we combine the outputs of four sensels to make a "fat sensel", and represent that not in 14 bits but rather in 15 bits. (Remember, if we add four values that are known to 14 significant bits each, the precision of the sum is (I think) 1 bit greater. I have to work on that a little!)

Click to expand...

Yes, averaging 4 sensels (for red or blue) will yield 1/Sqrt(4)=1/2 or 50% of the noise, and 8 sensels (for green) will yield 1/Sqrt(8)=1/2.82 or appox. 35% of the noise, so one could gain up to 1.5 bit of (green) accuracy and 1 bit of R/B accuracy, and DR. That is how it should be done, and the super-sampling should help with aliasing prevention of the half resolution as well. However, keeping in mind why the sRaw format exists in the first place, for speed reasons (and also with a 14-bit optimized ASIC / DIGIC) I could imagine just 1/4th of the sensels being read, quick-and-dirty.

I guess an analysis of a few Raw comparison shots is required, if we want to unravel this question, for the manufacturers are unlikely to tell us what happens under the hood. In theory, comparing the non-demosaiced read-noise or dark-noise in IRIS or similar applications should show the difference in noise, if any.

Bart

Hi, Bart,

Bart_van_der_Wolf said:

In theory, comparing the non-demosaiced read-noise or dark-noise in IRIS or similar applications should show the difference in noise, if any.

Click to expand...

Will the current IRIS .CR2 decoder decode sRaw files, or do we know when it will get that capability?

Best regards,

Doug

Doug_Kerr said:

Will the current IRIS .CR2 decoder decode sRaw files, or do we know when it will get that capability?

Click to expand...

I haven't seen any announcement of DCraw support for the Mark III's sRaw output, so I doubt if IRIS can handle it yet. But then we won't know for sure untill someone picks up the glove ...

We'd need a Raw and an sRaw darkframe (no lens with body-cap on, no noise-reduction, viewfinder covered, short exposure time), if possible simultaneously written to 2 separate cards. As soon as there is a true Raw converter that allows to examine non-demosaiced data from the sRaw, we're in business.

Bart

Doug_Kerr said:

Will the current IRIS .CR2 decoder decode sRaw files, or do we know when it will get that capability?

Click to expand...

That's highly unlikely. It doesn't officially recognize the regular mk3 RAWs yet, and loads the entire file, headers and all, as an image. sRAW from Canon is a new thing (as is full RGB pixels from Canon, if that's what sRAW is), so I doubt IRIS is looking for it. IRIS reads DNGs, though, so if the Adobe DNG converter can convert it, then IRIS should read that.

Do you know where any sRAW samples are?

Bart_van_der_Wolf said:

Yes, averaging 4 sensels (for red or blue) will yield 1/Sqrt(4)=1/2 or 50% of the noise, and 8 sensels (for green) will yield 1/Sqrt(8)=1/2.82 or appox. 35% of the noise, so one could gain up to 1.5 bit of (green) accuracy and 1 bit of R/B accuracy, and DR.

Click to expand...

From the perspective of the new pixel, perhaps. From the perspective of the *image*, the only real change is a possible loss of subject resolution, IMO. I have never seen anything better about a binned or downsampled image, compared to the original, when both are displayed at the same size.

The pixel-centric view of IQ does not lead to proper conclusions, IMO.

John Sheehy said:

The pixel-centric view of IQ does not lead to proper conclusions, IMO.

Click to expand...

Sensels are the building blocks of the image, they determine how good the demosaicing quality can be. If they individally suffer from noise, the demosaiced result will suffer from noise.

Resolution on the other hand is determined by the interaction of the MTFs of the system components MTFs, lens x AA-filter x micro-lens x sensel spacing. The latter (sampling density) can benefit from super-sampling.

Both the individual sensel quality, and the spatial interactions are vectors in total perceived quality, IMO.

Bart

Nu?

So, do we yet have any idea what the small raw output is really all about?

All we seem to know is that is has about 1/4 as many sensel reports as the regular raw output, and has a file size a bit smaller.

Best regards,

Doug