R3D Compression test examples , DSLR images thrown in as well

[Les Dittert]

I did some more extensive compression tests today and prepared some results for you to see here.
I chose a more difficult scene, a nature scene with lots of details, so this should be considered about as hard as it gets as far as compression goes.

First some tech details:
Lens: Tokina 11-60 as 16mm and f8 . Same lens on all frames here.
Focus: Even with f8 having a lot of depth, I racked focus on all images to make sure it had the sweet spot.
On the Scarlet I opened to f2.8 and used Magnify and Edge to center the focus rack range.
( It turns out that the depth of field was so deep that there really could not be focus errors anyway )
Cameras:
Scarlet set to ISO 250 and 1/500 sec 12 FPS. ( 12 FPS to allow RC3 as well) Shot 4096x2160 . RC3, RC6, RC8, RC11
Exposure did not clip, full histogram.
Canon XTi : My old $200 Craigs list camera. ISO 200. Same lens settings. 10 Megapixels, about same sensor photosite size as Scarlet.
Canon T2i : Same settings as above. Smaller photosites so I moved camera back to get the 1:1 pixel size close to the same.
( in other words, I wanted the 2Ti to expose 4096 pixels across the same scene as the Scarlet's 4096 . This allows close 1:1 pixel comparisons with no resize anywhere in the test images. )

Processing: debayer=full, RCX sharpness =1 , detail =2 , no denoise, fairly flat color saved to 16bit TIFF image sequences. I added some unsharp mask to these images in Fusion.
Before anyone asks, I did not use sharpness > 1 because I found that it tended to artifact and not add any real detail anyhow.
For the Canon CR2 RAW images, processed in Adobe Camera Raw and used zero sharpen and zero noise reduction. Saved TIFF.

The following are the images, with annotations above each image as needed. Pay attention to the tree trunk in the cropped images.
( I posted high quality JPG for these, as the quality=12 setting does not change the image as much as the compression changes happening in the underlying R3D anyways )


This is the only resized image, to show what the full frame view looks like:

Red, cropped in to show detail. RC3

Rc6 crop:

Rc8 crop:

And the RC11 crop:

In the next post I will continue with the DSLR versions.

 

[Les Dittert]

now the DSLR versions

now the DSLR versions

So these are the Canon images, as a reference to what uncompressed might look like.
Obviously you can't shoot as speed in Canon CR2 raw, but they are interesting to compare.


The Canon XTi , no resize, no sharpen in post:

The Canon T2i , note the tree trunk detail. This is roughly one to one size to keep the 4096 width , I moved back to do this.

And the Scarlet RC3 version to compare with:

In the next post, I add something interesting

 

[Les Dittert]

Adobe Jpeg 2000 versions !

Adobe Jpeg 2000 versions !

Here I take the 16 bit TIFF of the Canon images, and save Jpeg 2000 in Photoshop to get the same file size as the r3d snapshot from RCX. ( the RC3 version )

First the red RC3 image again:

And the Canon T2i saved at same file size:

I have the originals, etc, for those who want to explore them further.
It is obvious that RC3 is NOT LOSSLESS as some people seem to think. The quality setting that was used in the Adobe Jpeg2000 saver was something like 12 out of a possible range of 1 to 100.
Not a valid true comparison with Red Code, as I compressed an RGB image and RedCode compresses a RGGB color filter array image that has not been 'debayered' yet.
Just interesting to see.

 

[Adrian Jebef]

Great stuff Les. Demystification is helpful...

 

[Les Dittert]

For those who want to see the difference between Canon T2i and Scarlet RC3 in a more easy format, I blew them up 300% pixel replicats ( so you can see the pixels clearly, no fancy resize sharpen cheat here )
The images are color corrected to make it easy to see. Again, no rocket science here, one frame is compressed and the other isn't .
I have links for those who don't trust the Q=12 jpg images I'm posting here :
TIFF 16 bit lossless : https://dl.dropbox.com/u/61349012/DSLR_and_Red-compression_test/canon_T2i_Red_RC3_blowup.tif
Jpeg2000 16 bit lossless : https://dl.dropbox.com/u/61349012/DSLR_and_Red-compression_test/canon_T2i_Red_RC3_blowup.jpf

 

[Oliver Skibbe]

Les, your results reflect my own findings. Redcode is a very nice codec for shallow focus scenes. But it is unable to cope with some real world deep focus situations like wide-angle lenses at f8. Even at lowest compression resolution will drop below 3K.

I was hoping that Epic's high data rates would finally lift RED's "virtually lossless" compression into the "mathematically lossless" range, but unfortunately it did not. Maybe RED will give us some true lossless R3D mode for stills and time-lapse photography?

 

[John Marchant]

IMHO most of the difference you're seeing at RC3 is optical, not compression. RED's OLPF is pretty heavy, and the T2i looks digitally sharpened on top of a less aggressive stills-oriented OLPF.

No doubting the results at lower RC are heavily affected by compression in the shadows, but this is not news to me really...

 

[Glen Powell]

Did you add any post sharpening to the RED files to sort of match-ish sharpening added directly by the other cams?

 

[Evin Grant]

John is correct, this is not a compression difference, it's the OLPF or other optical limitation (diffraction maybe?). Each optical system is different and the Epic/Scarlet OLPF is tuned for cinematic "no aliasing at any cost" so it's reasonable to expect a slightly lower level of detail at 100%. The compression artifacts you see at 10:1 and 12:1 usually manifest as smudging of high frequency detail in shots like the ones above. Also compression artifacts are almost always more pronounced when the lots of high frequency detail is moving, like for instance a field of wheat in a breeze.

One interesting note is how much cleaner (color noise) the Scarlet frame is that the Canon at 300%

 

[Mike P.]

One interesting note is how much cleaner (color noise) the Scarlet frame is that the Canon at 300%

Yeah, but it's also noticeably softer. I'd imagine if the Canon shots were "smoothed" or "blurred" by an equally aggressive OLPF, the colour noise would be less noticeable. Conversely, I don't think post-sharpening would work at raising the Scarlet's detail level to the stills detail level, despite that the Canon's supposedly have post-sharpening always-on automatically.

 

[Les Dittert]

I did do some sharpening on the red frames, you can see that it effected the rocks edge. I could not get back the other details that are lost due to compression smoothing.
i could have noise reduced the canon frames but did not.

 

[Álex Montoya]

That's probably the most difficult scene for a Red camera. Everything's in focus and almost the full frame is filled with crazy detail.

In this scenario, of course, the compression starts to show, not so much in the softness which I agree comes more from the OLPF, but in things like the loss of nuances of the individual straws on the ground.

If you need more detail than this you should think about renting a F65 and shooting uncompressed.

Anyway, could you upload a single frame in R3D format? I would like to take a look at it.

 

[BRANDON JAMESON]

Here's another source of a frame grab from the new Aaton Delta Penelope to compare:

http://blog.abelcine.com/2012/09/27/aaton-unveils-the-delta-penelope-camera/

 

[Les Dittert]

It is a difficult frame, to be sure.

Here is the R3d and Rmd:

https://dl.dropbox.com/u/61349012/DSLR_and_Red-compression_test/A005_C001_0929SO.0000000_S000F.R3D
https://dl.dropbox.com/u/61349012/DSLR_and_Red-compression_test/A005_C001_0929SO.0000000_S000F.RMD

The r3d snap is 4.2 megabytes.

Some math: 4096 x 2160 x 2 = 17.69472 megabytes for an uncompressed 16bit image
This means that the compression ratio for the frame is 17.69472 / 4.2 = 4.21 ( close enough to the RC 3 setting )

That's probably the most difficult scene for a Red camera. Everything's in focus and almost the full frame is filled with crazy detail.

In this scenario, of course, the compression starts to show, not so much in the softness which I agree comes more from the OLPF, but in things like the loss of nuances of the individual straws on the ground.

If you need more detail than this you should think about renting a F65 and shooting uncompressed.

Anyway, could you upload a single frame in R3D format? I would like to take a look at it.

 

[Gabe Conroy]

It is a difficult frame, to be sure.

Here is the R3d and Rmd:

https://dl.dropbox.com/u/61349012/DSLR_and_Red-compression_test/A005_C001_0929SO.0000000_S000F.R3D
https://dl.dropbox.com/u/61349012/DSLR_and_Red-compression_test/A005_C001_0929SO.0000000_S000F.RMD

The r3d snap is 4.2 megabytes.

Some math: 4096 x 2160 x 2 = 17.69472 megabytes for an uncompressed 16bit image
This means that the compression ratio for the frame is 17.69472 / 4.2 = 4.21 ( close enough to the RC 3 setting )

That should be 16.875 MB...although that still comes out to a ratio of ~4. However since the ratio is meant to go as low as 3, isn't that telling us the image itself started off without the detail before it even got to compression? (e.g. due to the OLPF) Just like when I losslessly compress motion graphics and the compression ratio is ludicrously high because there's just not a lot of detail in the image...it doesn't mean the compression is being applied too heavily, it means the image lacks detail.

 

[Les Dittert]

That should be 16.875 MB...although that still comes out to a ratio of ~4. However since the ratio is meant to go as low as 3, isn't that telling us the image itself started off without the detail before it even got to compression? (e.g. due to the OLPF) Just like when I losslessly compress motion graphics and the compression ratio is ludicrously high because there's just not a lot of detail in the image...it doesn't mean the compression is being applied too heavily, it means the image lacks detail.

If the image lack detail, we have a bigger problem, because the image is very full of detail. If the OLPF cut the detail too much, that would be suboptimal.
Remember, I used same glass and careful focus on these tests. F8 is very deep focus as well.

As a data point, I compressed the Canon image ( cropped to 4096 x 2160 ) with Jpeg 2000 in Photoshop and got the 4 megabyte file size with a quality setting of 12 out of 100. I posted that image result as well. RGB compression is not directly comparable to CFA RGGB compression, but it's interesting to note how low I had to set the Q setting to get the small file size. 16bits used of-course as Red Code is claimed to be 16bits as well.

Where did you get the 16.875 MB number ?

 

[Gabe Conroy]

If the image lack detail, we have a bigger problem, because the image is very full of detail. If the OLPF cut the detail too much, that would be suboptimal.
Remember, I used same glass and careful focus on these tests. F8 is very deep focus as well.

Suboptimal for what? You're comparing a still camera to a motion camera. Still cameras don't need OLPF's as strong as motion cameras.

As a data point, I compressed the Canon image ( cropped to 4096 x 2160 ) with Jpeg 2000 in Photoshop and got the 4 megabyte file size with a quality setting of 12 out of 100. I posted that image result as well. RGB compression is not directly comparable to CFA RGGB compression, but it's interesting to note how low I had to set the Q setting to get the small file size. 16bits used of-course as Red Code is claimed to be 16bits as well.

Yeah but an RGB image is still going to be three times as big as a RAW image of the same resolution...I don't see how this tells us much of anything.

Where did you get the 16.875 MB number ?

1 kilobyte = 1024 bytes
1 megabyte = 1024 kilobytes

You used
1 kilobyte = 1000 bytes
1 megabyte = 1000 kilobytes

 

[Les Dittert]

I think I'd prefer a little less low pass on my camera, is what I think I'm concluding to after all of this.
The Black Magic camera chose to use none. I did see some aliasing on some sample DNG's but I'm sure certain textures will cause it a problem.
Can't wait to try it out.

The point with my J2K compress test in Adobe is, that even tho the RGB image is huge, as you say, three times bigger, the Adobe J2K compressor was able to reduce it to the same minuscule 4.2 MB frame size.
Even with that handicap, the sharpness exceeds the RC3 version. No OLPF is the main reason, seems like.
I know that the real time compression in the camera is taking some computational shortcuts to get the required performance, so that has to be considered. I did use the integer j2K option, to cripple the adobe compression just a bit.

I wonder how hard it is to remove and replace the OLPF?

Color aliasing can be improved upon a great deal with motion picture cameras.

Just for fun I'll rig the Canon T2i onto a motion control rig and simulate a motion shot of a similar daylight scene to see how it looks.
Come to think of it, Tom Lowe shot most of his time lapse movie "Timescapes" with a Canon still camera. I wonder how he feels about the aliasing, he certainly had a huge quantity of fine detail images there. Stars. Thin branches.
Perhaps he swapped in a heavier OLPF ?

 

[John Marchant]

OLPF is easy to remove, but its laminated with the IR cut filter, so a lighter / non-existant OLPF would need to be a specialist or RED-built product. (unless you want crazy coloured IR flooded pictures )

 

[Les Dittert]

Thanks for the info. Could always add IR cut to the front of lens for a test, I suppose.
Back focus would change too.

edit: Yep, I forgot that when I added the Red flare baffle , I actually had it removed and in my hand !
I peeked at the edge, and saw the sandwiched IR and OLPF. The whole thing is actually fairly thick, maybe 1/10 an inch.

I'll see if I can get hold of an IR cut filter and test ASAP.