Big New Canon Rumor - RAW Recording Module for 5D2 & 7D?

[Graeme Nattress]

Mine too... Cameras are complex beasts and lots of "gotchas" in their design.

Graeme

 

[Sanjin Jukic]

Jack, thanks for your thoughts on the technical feasibility of this. I have no way of knowing if this is true or not. I'm passing along this info "as is". On the other hand, I don't doubt that Canon engineers are capable of working miracles.

Regarding sRAW, I find this whole thing to be very interesting. I made a post in the DSMC forum back this winter, suggesting that Graeme and Co. consider offering an sRAW option on DSMC, to cut down on recording bandwidth. To make a long story short, Red had already looked at and considered an sRAW type of recording, but found problems with aliasing, artifacts, etc.

As I shot some timelapse tests with the 5D2 in sRAW 1 mode (half resolution) I eventually came to understand what Graeme and Jim and them were warning me about. :blushing: I started catching artifacts in my shots that would not have been there if I shot full RAW.

Here is an example:

This example is subtle, but it was disturbing to me. Notice the "ghosting" of those lights near the top of the image. I never get that on my 24mm 1.4 prime when shooting RAW, so I was concerned that sRAW itself might have caused it.

In addition (and even worse for me), sRAW "burns in" hot pixels from the sensor. In RAW mode, I can import RAW frames into After Effects and Adobe Camera RAW instantly kills all hot pixels. It can clearly identify them because it's looking at the entire sensor. But, when I shot sRAW 1, and imported to ACR, the hot pixels were totally burned into the "binned" image. The sRAW downsampling engine had assumed that the hot pixels were stars or bright lights, and combined them with surrounding pixels during the downsampling process. So now, ACR and AE were seeing the hot pixels as legitimate parts of the image, and did not remove them. It became very difficult to remove them. Another major factor that caused me to drop sRAW altogether.

Maybe if Canon used a smarter sRAW downsampling technique, hot pixels would not get burned in, so it's not hopeless in that regard at least.

Tom, I would say the "ghosting" issue is coming from the lens and not from sensor or RAW file.

Some super fast Canon and Leica lenses showing that sort of "ghosting" as usual.

Dots at the bottom are "ghosting" lens issue, shot on RED1 with Canon FD L 50mm f/1.2 @ wide open @ f/1.2.

Dots at the bottom are "ghosting" lens issue, shot on RED1 with Canon FD L 50mm f/1.2 @ wide open @ f/1.2.

Have a look those 'ghostings" much better at the end of this "available light" test footage shot hand-held on my RED1 and find some of similar "ghostings" made by Canon FD L 50mm f/1.2 @ wide open @ f/1.2.

Some older Leica Noctilux-M 50mm f/1.0 lenses made the same "ghosting" in my test on RED1.

LINK>>>

 

[David Newman]

Yes that Leica M8 has some ugly ghosting, and clearly seems to be a digital artifact. So it can happen. I'm just not dismissing the more likely scenario for the above image -- the reduction in scale of the ghost indicating a refection from a curved surface within lens. I have seen this, but then I don't always shoot with really nice lenses. If RAW is good, binned-RAW (sRAW) would have the same base image, things only going wrong at the detail level -- aliasing, dead pixel coupling, etc. If more can go wrong, resulting in the above image, I would love to know the mechanism how.

 

[Tom Lowe]

A look at sRAW zone plates would be one way of checking, I suppose.

 

[Jack Kelly]

...If RAW is good, binned-RAW (sRAW) would have the same base image, things only going wrong at the detail level -- aliasing, dead pixel coupling, etc...

Please could someone explain roughly how sRAW works? In particular, I don't really understand how it's possible to down-res a Bayer pattern image without first demosaicing it. i.e. I can understand how it is possible to produce, say, a 2k image from a 4k sensor by first demosaicing the 4k bayer pattern RAW to 4k RGB and then downsampling from the RGB. But how is it possible to pixel-bin a RAW image without breaking the Bayer pattern? Does it work by taking adjacent RGBG groups together? E.g. if the raw sensor output looks like this:

BGBGBGBGBGBGBGBGBGBGBGBGBGBG
GRGRGRGRGRGRGRGRGRGRGRGRGRGR
BGBGBGBGBGBGBGBGBGBGBGBGBGBG
GRGRGRGRGRGRGRGRGRGRGRGRGRGR
BGBGBGBGBGBGBGBGBGBGBGBGBGBG
GRGRGRGRGRGRGRGRGRGRGRGRGRGR
BGBGBGBGBGBGBGBGBGBGBGBGBGBG
GRGRGRGRGRGRGRGRGRGRGRGRGRGR

Then is one approach to downsampling the RAW signal to skip entire "RGBG" groups like this:

BG----BG----BG----BG----BG----BG
GR----GR----GR----GR----GR----GR
----------------------------------
----------------------------------
BG----BG----BG----BG----BG----BG
GR----GR----GR----GR----GR----GR
----------------------------------
----------------------------------

(i.e. the hyphens indicate sensor elements which are present on the sensor but are skipped during readout)

If it worked like that then I can kind of understand how it would be possible to pixel-bin the RAW Bayer patter image whilst still ensuring that it's possible to later interpolate the image to produce an image where each pixel has an RGB value without getting ugly colour aliasing artefacts. But I'm still a bit confused about how it's possible to down-sampling RAW without first debayering.

I suppose at the heart of my confusion is that I initially imagined that a pixel-binned RAW image might look something like this:

B--G--B--G--B--G--B--G--B--G--B--G--
---------------------------------------
---------------------------------------
R--G--R--G--R--G--R--G--R--G--R--G--
---------------------------------------
---------------------------------------

(i.e. skipping every 2 sensor elements)

But then wouldn't you loose the ability to accurately reconstruct an RGB image because the Bayer sensor elements are now quite far apart?

Sorry, this is a bit of a rambling post... just procrastinating because I've got a boring edit to do today!

 

[Matthew Bennett]

just procrastinating because I've got a boring edit to do today!

Aahh, the cry of the bored editor, with no other solace than checking reduser every five minutes...

I know it well,

And you get nothing but the sincerest sympathy from me...

god bless us all, and may some sort of interesting material happen to fall into our laps to cut...

 

[Graeme Nattress]

Looking at the file sizes of sRAW they bin to RGB, rather than keep a bayer pattern in the raw. I'll have to dig out the image I took in sRAW because it's using a binning rather than a proper downsample, the aliasing artifacts are increased and Canon have a pretty weak OLPF to begin with.

Graeme

 

[Stephen Williams]

There is a potential issue using lenses that cover FF35 on S35, lots of extra light bouncing about. Cooke make masks for use with the S4's when shooting S16 because of this.

Tom, I would say the "ghosting" issue is coming from the lens and not from sensor or RAW file.

Some super fast Canon and Leica lenses showing that sort of "ghosting" as usual.

Dots at the bottom are "ghosting" lens issue, shot on RED1 with Canon FD L 50mm f/1.2 @ wide open @ f/1.2.

Dots at the bottom are "ghosting" lens issue, shot on RED1 with Canon FD L 50mm f/1.2 @ wide open @ f/1.2.

Have a look those 'ghostings" much better at the end of this "available light" test footage shot hand-held on my RED1 and find some of similar "ghostings" made by Canon FD L 50mm f/1.2 @ wide open @ f/1.2.

Some older Leica Noctilux-M 50mm f/1.0 lenses made the same "ghosting" in my test on RED1.

LINK>>>

 

[Graeme Nattress]

This is a 1:1 grab of the zone plate, shot on a 1D mk III, in sRAW. Used Adobe Camera RAW to open it. You can see the vast amounts of aliasing, making for even the lowest frequency details to have a stair-stepped appearance. It's worth comparing to the over-all cleanliness of the same chart, same framing, on the RED One: http://www.cinematography.net/natress-red-res.html showing almost complete lack of such aliasing nasties.

Remember though, that while aliasing can be bad on a still, it's killer on motion as when things move, the aliases move in the other direction. This is a nightmare for any motion adaptive codec. Sure, you mightn't be shooting with a motion adaptive codec, but you surely will be delivering on one for broadcast TV, and downscaling only ever makes aliasing worse, not better.

Graeme

 

[Graeme Nattress]

Is same chart in RAW. Ouch. Although this time converted in Apple's Aperture.

Graeme

 

[Tom Lowe]

Wait... are you saying that RAW is worse? I would stick to ACR or DPP.

Would Panavision in Hollywood have good zone plates? I'd like to shoot some 5D2 sRAW1, sRAW 2 and RAW frames, maybe even with camera timelapse motion, and see how they hold up. Then again, if this external recorder is for real, maybe I will wait for that to shoot these tests, because the sRAW on that recorder might be a different flavor than the stills method. For one, I assume they would come up with some way to kill off hot/dead pixels before binning. Of course, this is assuming that the Canon sRAW recorder is real.

 

[Roberto Lequeux]

Graeme, wouldn't we have to see a full resolution version to see the lack or amount of aliasing? Or were those 1:1?

 

[Graeme Nattress]

This is from Adobe Camera Raw.

 

[Graeme Nattress]

And for completeness, here's Canon DPP.

Graeme

 

[Graeme Nattress]

All those were 1:1 crops. Aliasing across RGB is pretty constant in sRAW and that's to be expected with the binning. In RAW, luma aliasing is not too bad, and certainly a lower level than that of sRAW. Chroma aliasing is not handled well though. However, usually, raw developers should be able to handle that quite well.

And because demosaic algorithm effects the results, I've just posted results of the same file developed in ACR and DPP. All show similar, but different demosaicing issues. I think the DPP problems are made worse by the default tone curve, which seems more aggressive than Apple's or Adobe's. Looking again, maybe it's more like default sharpening in DPP.

Graeme

 

[Roberto Lequeux]

So the post #73 seems really clean... other than all that horrid color funk!! What the hell is that?

 

[Graeme Nattress]

Yup, Adobe doesn't do too bad a job on it at all. It's pretty clean in the luma, as I said - that's not too bad. But in the highest frequency areas, at details above which the camera should not be able to see (that's a 4k zone plate on a camera with 3888 pixels horizontally), the false aliased detail gets interpolated as funky colors by the demosaic algorithm. That was shot with the 50mm f.14 at f4, so pretty darn sharp, so pretty much worst case.

With sRAW you're trading lack of chroma nasties for lower overall resolution and increased luma aliasing. On real world images, you're more likely to see luma aliasing than chroma aliasing as usually real world things have more visible detail in luma than in chroma, and usually luma edges are more contrasty than chroma edges.

Graeme

 

[GL2]

Canonrumors.com just posted a very interesting email from someone with insider info from Canon. For those who don't know, DIGIC IV is the current-generation processor used in Canon's cameras. The just-announced 7D will have two. DIGIC V--you guessed it--is the next generation processor.

I'll save you the click...


5D2, 7D and 1D4
Digic IV was not designed as a video processing chip, hence it cannot downscale multi Megapixel images in real-time at a frame rate needed for video. The fastest camera ever released by the company is the 7D and it can only down sample by pixel binning close to ten 18MP images to any smaller size a second. Current implementation skips vertical and horizontal lines then averages pixel values.

Digic V has a new on die area for video processing which lets it do up to 60 1080p frames a second with help from a pre-encoding ASIC chip that does pixel binning really fast. This will generate an image with almost no visible artifacts or moire. It will still be a step below RAW video but quality will be usable on production environments and on heavily detailed scenes.

I have not heard anything about a RAW module for SLR cameras. It seems far fetched since the CF module on these cameras just can’t feed that kind of data for video even in sRAW. The only raw module I have seen specs on will be an option package on the upcoming line of video specific products in 2010.

The 5D MII 24p firmware seems to be on hold, it has been tested since the camera was a prototype and there is simply no reason why the 5D can’t do 1080 24p or 720p 30. The upcoming 1D Mark IV will have even better video than the 7D from the specs Ive heard about and this camera will not be released this year as speculated. So Canon doesn’t
really need to release a new update for the 5D. One interesting
thing that might cement this decision is that the rolling shutter on the 5D MII is much nastier in 24p and they probably thought it was just not good enough to release it. If the patch is released expect heavy rolling shutter.

1D Mark IV what to look for based on part lists from suppliers and engineers:

-Completely new AF module akin to that on the 7D but with several times as many AF points. The tracking performance on this will be the best in the industry for many years. Video focus assist will be improved in speed five fold.

-Video: 1080p 24,30,25 720p 50, 60, SD with improved picture and audio control limits. Improved video processing. If it has Digic V it will be a true binning downscale and a big jump in quality.

-No swivel screen, improved contrast and color gamut.

-12 fps at full 14bit. – this is not confirmed just rumored by staff outside of the project!

The 1D line is guarded air tight and even many engineers working on modules for the camera have not even seen the full spec list.

http://www.canonrumors.com/

 

[Roberto Lequeux]

Wow. It reads believable.

But the 7D sounds like a decent buy. The $4,000 price of the 1D is still significantly higher than the 7D's, and the replacement for the 7D wouldn't happen till late 2010 at least, right?

On the RAW side, Scarlet Fixed would still blow away the 1D. Unless Canon is also coming up with a good "pro" RAW solution. No mention of that but it could be part of the 1D program. But for "pro-sumer projects" the 1D sounds like it might be a great way to go if there really won't be image monsters peaking their heads and ruining shots. Unless you need slow motion, then forget it.

However, what's interesting is that these technologies are supposed to be shared with their video camera department.

Sounds like a drag race where Canon's driver has some serious reaction time issues. Things really got stirred up, huh?

Good thing Jim has some experience with drag racing.

 

[GL2]

the replacement for the 7D wouldn't happen till late 2010 at least, right?

We don't know how long the product cycle will be for the 7D. My guess is 18 months, which means we won't see a 7D mark II (no doubt with a DIGIC V processor) until early 2011.

On the RAW side, Scarlet Fixed would still blow away the 1D. Unless Canon is also coming up with a good "pro" RAW solution.

I thought this quote from the email was particularly interesting:

"The only raw module I have seen specs on will be an option package on the upcoming line of video specific products in 2010."

Sounds like the sleeping giant may be waking...