Split Compression

[Guy]

After reading this post:

http://www.reduser.net/forum/showpost.php?p=79549&postcount=38

I took a closer look at this image(48MB TIFF):

http://www.hdforindies.com/uploaded_images/B016_C001_070902_001_OCN.tif

After messing with levels and looking at the blue and green channels by themselves there is a noticeable change in the texture of noise in the sky between the top half and bottom half of the picture.

The change happens exactly(to the pixel) in the middle of the picture and is visible on both sides of the car.

To me it looks like that the top half of the image and the bottom half are being compressed independently of each other. It is probably much more apparent on images like this one where the top half and bottom half are very different.

I wonder if this is baked into the hardware of if it is something that could be changed with different firmware.

Graeme?

[Michele Gavazzeni]

Probably the amount of data generated by the sensor exceed the bandwidth available from a single processor
in these case it could be handled by two separate chips one for the top half one for the bottom half.

In these case something isn't working properly giving different results.

it looks like the image has been processed with a soft median filter

Maybe an adaptive noise reduction filter is applied when capturing
so if like in these shot:
the top part is mainly sky
while in the bottom half has more elements with different colors and noise patters.
In these case an adaptive filter would have read two different scenarios applinig different corrections!

This joust an idea i've no knowledge about how Redone really works.

[Ryan Sims]

Here's my guess: The top half of the frame is mostly blue with very little detail. The blue channel got most of the bandwidth there, hence more detail in the blue channel as seen as grain. The bottom half has very little blue and lots of detail, so the red and green channel got nearly all the compression bandwidth. Blue didn't need much info for the bottom. Good info to know if you plan to do green or blue screen VFX work. Up until last week, I think most of us had never worked with 4K moving picture files (even though I'm just playing with files you guys have posted). RED truely is 4K for the masses!

Here's the images we're talking about:
http://i111.photobucket.com/albums/n134/im_thatoneguy/RED_Compression_tile.jpg

http://www.reduser.net/forum/uploaded/697_1189372756.png

That came from the uncompressed file of this:
http://www.hdforindies.com/uploaded_images/B016_C001_070902_001_OCN-797507.jpg

It would be best to also read this whole thead as well:
http://www.reduser.net/forum/showthread.php?t=4330&page=5

[Miltos Pilalitos]

I have noticed the same thing way back when the first frames from "Crossing the line" were posted. I didn't want to mention it back then because it was from a beta camera.

I don't think the reason is that one particular channel is taking all the bandwidth. You can see the same effect in the following frame even if you don't watch a particular color channel. Observe well...
http://www.reduser.net/forum/uploaded/534_1189387100.jpg

For a full frame: http://www.red.com/skin/img/gallery-still/009000.jpg

the split is exactly on the middle of the frame!

[Luis Otero]

De-Ja-Vu from JVC Split screen issue?

[Ryan Sims]

Here's a split example in the City file as well:
http://www.reduser.net/forum/uploaded/697_1189389307.png

You have to tweek the contrast really hard to see it.

[Miltos Pilalitos]

De-Ja-Vu from JVC Split screen issue?

:biggrin: Are you suggesting that Mysterium's secret is that it's actually 2 CMOS Chips? :biggrin:

[Luis Otero]

No, I am not suggesting anything. It is just a surprise issue that we went through JVC HD camera... The technical explanation must be given by RED, and I am not associated with them at all (just RED resrvation # 711).

[Guy]

You can also see it here in the blue channel of the Garage TIFF(this is without any kind of manipulation), 100% crop:

[Miltos Pilalitos]

Hmm.. interesting...

Before you post this i thought that the upper part was always the grainy one and the lower the softer but in that pic it's seems that the contrary is also possible.

Strange!

[Steve Freebairn]

I guess for those that really are going to use this as a replacement for their DSLR will have to use the raw port. I'm curious to see if this issue is something that will go away with improved firmwares? It seems like if they are use two separate processors to process the data, they could "talk" to each other and make the line go away. Maybe there was a hidden message in Crossing the Line (sorry I couldn't resist, I'm personally thrilled by the images I've seen so far).

[Joel Kaye]

I'm curious to see if this issue is something that will go away with improved firmwares?

Compression can be changed in firmware, so yes it could. Personally, I think this is so minor I'll never notice it if it never changes.

[Ryan Sims]

After looking very hard, the split can be seen here slightly in the Skytower shot (this is 100% zoom and the contrast and brightness really cranked.) Interesting because the the city shot is from the same camera under similar lighting conditions.

http://www.reduser.net/forum/uploaded/697_1189394650.png

This is nothing when compared to the original JVC split processing problems. I'm not worried. In fact, I'm pretty happy with the image I get on my Canon XH-A1 most of the time. RED's image is far superior to it.

[Ryan Sims]

Ok, Look at this blue channel picture from Evin's post today of the updated PSgarage. This looks nothing like the what the old TIFF did here. It may be a slightly different frame, but it still has none on the split problems. I'm beginning to believe that you can't really trust uncompressed TIFFs or DPXs at all. What matters is only what you do in REDalert/cine with the raw files first.

http://www.reduser.net/forum/uploaded/697_1189397757.png

Guy, I assume that you were working on the TIFF from last week. Not PSGarage. I can tweak the old file to look like that as well.

[Joel Kaye]

In the latest fxphd podcast which covers RED they warn against judging posted images because it's like judging post telecine. It's not a true raw evaluation. They own camera 22. I don't know if that impacts this issue but keep it in mind.

[Ryan Sims]

Thanks Joelnet for the update. Good news everybody! Forget about downloading the massive TIFFs and DPXs. Wait and download REDCODE RAW files later when we have a copy of REDcine. For now, the JPEGs will do just fine. The TIFFs are just causing confusion.

It would still be nice to keep posting some of your best shots as uncompressed 4k as well.
4K PNGs would be great. See my other thread on PNGs.
http://www.reduser.net/forum/showthread.php?t=4487

[Joel Kaye]

Thanks Joelnet for the update. Good news everybody! Forget about downloading the massive TIFFs and DPXs. Wait and download REDCODE RAW files later when we have a copy of REDcine. For now, the JPEGs will do just fine. The TIFFs are just causing confusion.

It would still be nice to keep posting some of your best shots as uncompressed 4k as well.

Is there a full moon tonight?

[Miltos Pilalitos]

RedRyan i checked almost everything i have of stills and footage from the camera and i noticed that the split compression issue is mostly apparent with the JPEG captures and not in all of them.

Maybe the problem is not with the footage but with the way REDAlert exports the 4K jpegs.

I hope someone from RED will drop a line to illuminate the subject.

[Guy]

Guy, I assume that you were working on the TIFF from last week. Not PSGarage. I can tweak the old file to look like that as well.

Yeah you are right, weird. Wonder why they are different?

[Álex Montoya]

There's also split compression (much more noticeable than these examples, so they have improved a lot) in the Crossing The Line JPG frame grabs.

Check the one with the soldier in the ground.

[Gavin Greenwalt]

It's not a jpeg issue. It's in the uncompressed. I would be 99% confident that it's a codec 'feature'.

Jpeg2k uses multiple tiles which become apparent in noisy images. I was hoping RED wouldn't employ tiles but so be it, hopefully once CF cards increase in speed we can disable that 'feature'.

[Justin Anderson]

It's not a jpeg issue. It's in the uncompressed. I would be 99% confident that it's a codec 'feature'.

Jpeg2k uses multiple tiles which become apparent in noisy images. I was hoping RED wouldn't employ tiles but so be it, hopefully once CF cards increase in speed we can disable that 'feature'.

What if you just split the image into more tiles? Shouldn't that reduce the noise variation?

[Jannard]

This is a known issue with a fix on the way. It will be a simple firmware upgrade. We had it with Boris and Natasha, thought we fixed it and it resurfaced recently. We are on it.

Jim

[erbp]

This is a known issue with a fix on the way. It will be a simple firmware upgrade. We had it with Boris and Natasha, thought we fixed it and it resurfaced recently. We are on it.

Jim

Thanks for clarifying. Will the compression artifacts in the blue channel be fixed as well?

[Álex Montoya]

Nice, Jim.

[Steve Sherrick]

Thanks for clarifying. Will the compression artifacts in the blue channel be fixed as well?
Yeah, blue channel is really suffering. I certainly see some compression in the other channels as well, but not nearly as much.

Steve

[Graeme Nattress]

We're continually working to improve the compression in all it's aspects.

Graeme

[Steve Sherrick]

No doubt Graeme. This has to be a huge challenge trying to maintain high level of quality, yet keeping the size manageable for the format. Is it possible that we're seeing any weirdness introduced in these TIFFs? I'm really looking forward to seeing footage come directly off the camera to make a more hands on judgement.

Steve

[Graeme Nattress]

Yes, it's a challenge. It's doable, but a challenge. Each and every image you put through a scheme is different, and compression will act in different ways. As Jim mentioned, we saw this early on in B & N, but we put some more code in to stop it happening. That it's come back is annoying, but there's still plenty of things left for use to try!

Graeme

[Steve Freebairn]

Yes, it's a challenge. It's doable, but a challenge. Each and every image you put through a scheme is different, and compression will act in different ways. As Jim mentioned, we saw this early on in B & N, but we put some more code in to stop it happening. That it's come back is annoying, but there's still plenty of things left for use to try!

Graeme

Great news!, I'm sure you'll get it knocked out. So far, even with the minor issues, the footage still looks great.

[Steve Sherrick]

Graeme, I assume you have played around with some footage from the RAW port. Based on how good the images look in Redcode RAW even with some of these issues popping up, I imagine the RAW stuff must look even more breathtaking. However, from the beginning you guys have said it was hard for people to tell the difference when playing them back to back. Are these artifacts just tougher to see when playing back the 4K in motion on high quality monitor or projector? Are these things that people are just not catching until you look really carefully at stills? The footage I have downloaded that has been in motion has been .264 and any banding or other compression issues I have attributed to that codec.

Sorry for all of the questions. A lot of my curiosity about compression is directly related to a potential client's needs to do exteremly high end compositing and I am concerned about any issues that would give them reason to not use the camera.

Thanks,
Steve

[Nik Manning]

All I can say is I would have never saw that line and still don't really see it. If any one wants to switch my hv20 for their Red until they figure out this deal breaking problem I am right here for you buddy. :)

[conrad gaunt]

I have noticed the same thing way back when the first frames from "Crossing the line" were posted. I didn't want to mention it back then because it was from a beta camera.

I don't think the reason is that one particular channel is taking all the bandwidth. You can see the same effect in the following frame even if you don't watch a particular color channel. Observe well...

For a full frame: http://www.red.com/skin/img/gallery-still/009000.jpg

the split is exactly on the middle of the frame!

is that a wavelet based jpeg2000 or a normal 4:2:2 jpeg recompress? If it 4:2:2 it will screw with the blue, and the red a little too

[MikeCurtis]

I just LOVE the development process of this camera!

It used to be that a camera shipped, that's probably just how it was going to be, forever more. Minor, minor things could be tweaked with a firmware update, but it was often a fairly big deal to do so.

No more!

Get an email, load up a CF (soon SD) card with the fix, and you're off to the races with new/improved functionality.

In the 10 days I was in NYC with Offhollywood, I saw 2 firmware updates come in and multiple code builds as things ever improve. Not just fix, but IMPROVE.

I love it!

Even as software dev goes, these guys are INCREDIBLY fast. Forget hardware vendors - "Uhh...check back with us at NAB."

-mike

[jbeale]

A lot of my curiosity about compression is directly related to a potential client's needs to do exteremly high end compositing and I am concerned about any issues that would give them reason to not use the camera.
If you're at liberty to say, do you know what camera the high-end compositor would use instead? The Red compression issues I have seen so far are quite small in comparison to normal film grain, unless maybe you are shooting 65mm film or IMAX (?)

[Rob Lohman]

is that a wavelet based jpeg2000 or a normal 4:2:2 jpeg recompress? If it 4:2:2 it will screw with the blue, and the red a little too

We don't do 4:2:2

[Evan Owen]

We don't do 4:2:2

Not in the strict sense of the term, no. But all the shots I've seen show lots of resolution in the green channel, which doesn't seem to have any compression artifacts (visible anyway), while the red and blue channels exhibit significant compression artifacts and loss of detail.

At 4K, it doesn't seem to affect the image quality as much as 4:2:2 would normally, but you have to admit, there's not exactly an equal level of resolution among all three channels.

[Rob Powell]

If you're at liberty to say, do you know what camera the high-end compositor would use instead? The Red compression issues I have seen so far are quite small in comparison to normal film grain, unless maybe you are shooting 65mm film or IMAX (?)

I have similar concerns due to our compositing applications. It's not the compression that is the issue, per se. It's the fact that it is split. Trying to composit an image with settings that don't work for the whole image is problematic.

You're right, film does have grain, but at least it's consistent. While this artifact is not as noticable to the eye, it is exactly the type of difference that software could potentially lock on to creating varying noise patterns in a composited shot.

[Cail Young]

I have similar concerns due to our compositing applications. It's not the compression that is the issue, per se. It's the fact that it is split. Trying to composit an image with settings that don't work for the whole image is problematic.

As long as the compositor knows where the line is I don't see any problem - one already deals with spill suppression, holdout mattes, garbage mattes... a simple geometric mask shouldn't be too hard, but as I said, it means the compositor needs to know exactly where that line is.

[Graeme Nattress]

Sure uncompressed looks "better", but actually, the method of RAW decode / demosaic actually makes more difference to the end quality, so although I'm not using pristine Frankie uncompressed footage, the current compressed footage actually looks better.

Graeme

[Evan Owen]

As long as the compositor knows where the line is I don't see any problem - one already deals with spill suppression, holdout mattes, garbage mattes... a simple geometric mask shouldn't be too hard, but as I said, it means the compositor needs to know exactly where that line is.

Except that it's not just a simple line. Which side of the line was compressed more differs from shot to shot, and there are often other definite edges showing differing levels of compression throughout the frame. Manny's shots, especially girl_02.png, show it pretty clearly.

[Rob Powell]

As long as the compositor knows where the line is I don't see any problem - one already deals with spill suppression, holdout mattes, garbage mattes... a simple geometric mask shouldn't be too hard, but as I said, it means the compositor needs to know exactly where that line is.

You're right, but IMO it's tough enuf to deal with the "normal" challenges of compositing, without having a known issue that effects every frame and requires an addtional matte or adjustment.

[Rob Lohman]

Not in the strict sense of the term, no. But all the shots I've seen show lots of resolution in the green channel, which doesn't seem to have any compression artifacts (visible anyway), while the red and blue channels exhibit significant compression artifacts and loss of detail.

At 4K, it doesn't seem to affect the image quality as much as 4:2:2 would normally, but you have to admit, there's not exactly an equal level of resolution among all three channels.

Of course, bayer has two green channels and "only" one red & green. Since that is an important fact in a debayer algorithm you focus on that, obviously.

4:2:2 implies you actually drop full samples while they were available. After 4:2:2 there is still compression happening in those systems. We have the compression step, but don't drop samples before it.

[Evan Owen]

Of course, bayer has two green channels and "only" one red & green. Since that is an important fact in a debayer algorithm you focus on that, obviously.

4:2:2 implies you actually drop full samples while they were available. After 4:2:2 there is still compression happening in those systems. We have the compression step, but don't drop samples before it.

Right.

One thing I've been wondering. If the compression takes place on the pre-demosaiced RAW data, why is the green channel so pristine compared to the other two? Yes, I know there are twice as many green photosites, but that would mean the red and blue channels would show only about twice the compression. It looks to me like the green channel hasn't been compressed at all compared to the other two.

This whole 'compressing RAW' thing gets confusing to me sometimes. :blink:

[erbp]

I am interested in the science of what's happening, but mainly I would appreciate an answer about the artifacts along the line of "this is the best (or nearly) that can be done given the bandwidth of Compact Flash" or "we think we can get the blue channel as pristine as the green".

Also, I understand that we are talking about something that an audience watching on any modern presentation (less than 4k) isn't going to notice. However, it is affecting overall quality. I would not have even thought to look at the different channels if there wasn't a noise problem apparent in the RGB picture.

[Evan Owen]

...
I would not have even thought to look at the different channels if there wasn't a noise problem apparent in the RGB picture.

I wouldn't exactly call it a 'noise problem'. RED's sensor has the lowest noise of any digital sensor I've ever seen... :usd:

I'm talking about this stuff alot because I'm interested in the science behind it, not because I don't like the image and feel a need to complain about it... but yes, it'd be nice to know if (like I suspect) these things will be improving in the future.

[Stuart English]

Guys, we absolutely appreciate the interest in this subject. But its also only fair to say the precise details of what we are doing is RED team internal information only. So forgive us if we don't answer your specific points.

The most important information is: we are never satisfied, and will always look for ways to improve.

[erbp]

I wouldn't exactly call it a 'noise problem'. RED's sensor has the lowest noise of any digital sensor I've ever seen... :usd:

I'm talking about this stuff alot because I'm interested in the science behind it, not because I don't like the image and feel a need to complain about it... but yes, it'd be nice to know if (like I suspect) these things will be improving in the future.

I didn't mean anything I said as a complaint. And you're right, maybe "noise problem" isn't the right phrase. Here's where I'm coming from: I didn't reserve a Red to replace an HVX or any other digital sensor. I did it because I'm interested in replacing the workflow and reliance on labs that goes with 35mm. Right now, in my opinion 35mm beats Red as far as latitude and resolution but just barely. I would still choose Red because of the workflow, but I sure would feel better if I didn't feel like I was trading quality for ease of use. And if this compression can be improved that eliminates 35mm so I don't have to even think about it.

[jbeale]

Right now, in my opinion 35mm beats Red as far as latitude and resolution but just barely.

Have you done a controlled test (Red & 35mm side-by-side, shooting same scene at the same time with the same lenses)? With what film stock? I would be very interested to hear the detailed observations of anyone who had done this.

[erbp]

Have you done a controlled test (Red & 35mm side-by-side, shooting same scene at the same time with the same lenses)? With what film stock? I would be very interested to hear the detailed observations of anyone who had done this.

No, I haven't done a test. I don't get my Red until December. I would also like to see a side-by-side.

It's just an opinion based on feature work with Kodak 5217 and 5201 and scouring the forum for every image that's posted. I'd love to have access to RedAlert to see the latitude for myself.

[conrad gaunt]

We don't do 4:2:2

No, but people post jpegs images on the forum, which might make images look worse than they are? Trying to help you out here...

[Häakon]

Have you done a controlled test (Red & 35mm side-by-side, shooting same scene at the same time with the same lenses)? With what film stock? I would be very interested to hear the detailed observations of anyone who had done this.
I have, and I will be posting observations soon. I'm actually in talks with a fellow RED User who has been kind enough to offer his suite so we can all look at the footage on a 2K projector next to a 35mm projector at the same time. I've already done enough examining on this footage to know which I believe will look better, but I would like to see them projected simultaneously to really see for sure.

Film may still have a slight edge in the latitude department, but I'm not convinced it holds more resolution. At least not 35mm. I think if you saw them both projected in their full size you would be pretty blown away.

[Gavin Greenwalt]

Although I have to say on the other side of the coin--having just watched cloverfield: 720p looked fine by my eyes.

If you're going to a film print, I think George Lucas might be right: 1080p is more than enough.