Noise: shrinking High-MP vs Native Low-MP

[dr croubie]

So there's bsically 2 camps in the "what I want from the 5D3" various threads:
The High-MP camp, looking for something to beat the (alleged, i'll believe it when i see it) 36MP D800. Beat as is beat it in MP.
The High-ISO camp, looking for a see-in-the-dark camera, to beat the current lord-of-darkness D3S.

Now, If you take a low-MP camera, and upscale the image to high-mp, you don't get nearly as much detail for printing billboards (sometimes I think everyone prints billboards at 300dpi on a daily basis given what I read here, yet noone gets paid enough for these billboards to hire a Phase One IQ 180).
But, if you take a high-MP camera, you can shrink it down to the same printable-size as the low-MP camera, but is the noise better or worse?

Here's my basic test I just did. With a 7D and Samyang 35/1.4, taking a shot of my garden at various ISOs, in both RAW and sRAW.
Conveniently, sRAW on a 7D is exactly half the width and half the height of a RAW file, so 1/4 the size at 4.5MP. That means that every pixel is made up of one block of RGGB photosites (which makes 4 pixels in RAW mode). Now I don't know exactly how canon makes the sRAW file, but my guess is that shooting in sRAW will equate to what you'd get if they made a native 4.5MP sensor using the same sensor tech available at the time (it is for this reason you can't compare, say, the 1Ds2 to shrunk-1Ds3, because the processes and designs are 3-4 years newer).

Is it better to shoot in full-size RAW, then shrink when printing, or to shoot in the sRAW size that you need?
All shots are DPP processed, Standard style, WB cloudy, Sharpness 3, NR 0:0, ALO, PIC, etc all off, and converted to 100% jpg.
The 4.5MP sRAW shots were cropped at 500x500 and saved as 100% jpg.
The 18MP RAW shots were cropped at 1000x1000 (for same framing), then shrunk to 500x500 (which is what would happen if you print it on the same size paper as the sRAW shot), using GIMP, then saved as 100% jpg.

The 4 shots i'm posting below are at iso 3200 and 6400, purposefully to make noise as bad as possible to show up better. (Look at the filenames to see which is which).

 

[dr croubie]

So what's better?
For my money, the 18MP shots both show less noise once shrunk to 50% than the 4.5MP shots do.
So does that mean a 36MP sensor (using the same processes, and interwiring-to-photosite-size ratio as the 1DX) in the 5D3, and printed at the same size as a 1DX, will show less noise? Maybe, there might be other factors to take into account (like a square-proportional instead of linear-proportional response to shrinking photosites). I'm an electronic engineer, but i'm not a sensor tech.

Granted, with this test, the sRAW has 4 lots of wiring for effectively 1 pixel. A native 4.5MP sensor built using 7D tech will have only 1 wire per pixel, but then it will only capture 1 colour in the bayer-filter. So with a native 4.5MP sensor, the noise will be less than 7D's sRAW, but 2 of 3 colours' information will be not as good? How big is the interwiring compared to a photosite size? If it's 1%, then I can't see that making much difference. If it's 50%, then reducing MP count will have a bit more of an impact to noise...

But for sure, if I need to take photos in low-light, with one camera, I'm sticking with RAW then shrinking in post, rather than shooting in sRAW (which I may have tried in the past). mRAW and sRAW are now only for card-space saving, not better-noise.

That was only a basic test, if anyone has any better ideas for tests (without waiting for DXOMark to test sRAW), I'd love to hear them.

(you may now commence the arguements)

 

[dr croubie]

And another comparison I just thought of, just in case the 50% shrinking in GIMP wasn't the best.

First shot is the 4.5MP sRAW shot, at 500x500 pixels cropped as before.
Second shot is the 18MP RAW file, but DPP-processed to 2592x1728 in the "convert-and-save" dialogue box, then a 500x500 crop taken.
Again, 0:0 Noise reduction on both, and again the shrunk-18MP RAW file wins (in my eyes at least).

 

[Loswr]

I think the DPP processing looks better than GIMP. But I'm not sure the comparison is entirely valid. The in-camera conversion to sRAW isn't just downsampling the image. As the image is downsampled, it's increased from 14- to 15-bit to prevent quantization error, then 4:2:2 chroma subsampled (i.e., color detail discarded). That's done because just downsampling isn't enough file size savings for sRAW - before downsampling, the file must be demosaiced (in camera), so even downsampled, without the chroma subsampling it would be larger than the original RAW file

 

[pedro]

intresting post. while the rumour mill is gaining speed over at Nikon we hopefully may expect an intresting spring in Canonland.. Let's hope for 18 MP D800 and then wait what Canon come out with 8)

 

[NormanBates]

you need to compare two different sensors

or just go to dxomark.com, which does exactly this: resample all images to same target resolution, print them, and look at resolved detail and noise in the printed image

 

[lol]

Interesting test, but with too many unknowns to draw a conclusion.

On the "similar sensor tech" question, I have to wonder if the 15MP sensor from 50D and the 21MP from 5D2 are near enough the same sensor tech generation, since they came out about the same time and to me seem to have similar characteristics at relatively high ISO for each. So I propose crop the 5D2 output to APS-C size (~8MP) and see how that compares to the 50D resampled to same image pixel dimensions?

 

[awinphoto]

I think the DPP processing looks better than GIMP. But I'm not sure the comparison is entirely valid. The in-camera conversion to sRAW isn't just downsampling the image. As the image is downsampled, it's increased from 14- to 15-bit to prevent quantization error, then 4:2:2 chroma subsampled (i.e., color detail discarded). That's done because just downsampling isn't enough file size savings for sRAW - before downsampling, the file must be demosaiced (in camera), so even downsampled, without the chroma subsampling it would be larger than the original RAW file

That's exactly what I thought... You need to seperate cameras/sensors shot at full res and then downsample so they match... maybe a 7d vs rebel? Well maybe a lower res rebel...

 

[dr croubie]

Yeah, it's not the best test I know (the scientist in me is not too happy).

Comparing different sensor is what I was trying to stay away from, because in the years between their manufacture other processes would have changed (think of how your intel cpu is now made using 32nm-sized gates, 10 years ago they were 300nm). smaller interwiring means proportionately more space for photosites, as well as lower power reduces self-heating-noise.

But yeah, comparing sRAW to RAW might not be that good either, I was making a big assumption that 4x rggb photosites are just binned into one pixel at 1/4 the gain each, that may or may not be the case. And there's 4x the interwiring for that one pixel anyway.

The 50D vs 5D2 might have merit, if they're made using the same tech (i'm not awake yet so i can't think about the aps-c vs ff size factor though).
How about the 1000/1100D? Might they be made using the same sensor tech as current higher-MP models? In that case, they should have better low-light than a camera released at the same time? According to dxomark, the 10MP 1000D is just as good as the 15MP 50D, and the 12MP 1100D is below the 18MP 600D.
Curiously, the 550/600/60D are all the same, and a step below the 7D. Are they the same sensor really? Are they 'binned' for noise performance to determine what camera they go in (like most cpus are binned for speed for labelling)? Or are canon artificially crippling their noise performance somehow? (a 12 MP $400 1100D beating a $1500 18MP 7D wouldn't be good marketing, would it?) (But then, why wouldn't they cripple the 5D2 to be worse than the 1Ds3?).

Basically, all I wanted to do was to add some actual facts to the "i want 36MP 5D3" vs "I want 18MP see-in-the-dark 5D3". No matter what they release, half the people on this board are going to complain.
Could this be another 'split' in the 5D line, besides video? Like the nikon D3S/D3X low-mp-high-iso / high-mp-lower-iso split?
They've just recombined the 1D/1Ds lines into the 1Dx for shutter speeds, wildlife/birders aren't too happy because now they have less pixels-on-target than the 1D4. Could they now introduce a 36MP 2/3/4D studio camera (to replace the 1Ds for high-MP), and a 5D3 at 18MP (1DX sensor) for see-in-the-dark (weddings/concerts/anything)? Then the birders can use the 2/3/4D and crop happily, as long as it inherits at least the old 1Ds 45pt AF and 6+fps...

 

[Loswr]

I was making a big assumption that 4x rggb photosites are just binned into one pixel at 1/4 the gain each, that may or may not be the case.

No, they don't just bin - that would have significant negative consequences for a traditional sensor design like Canon uses. Check out the rearrangements, etc., that Fuji (their EXR sensors) and Phase One (e.g. P65+) had to incorporate to make binning work with a Bayer mask. For a traditional sensor, the image must be demosaiced first.

Curiously, the 550/600/60D are all the same, and a step below the 7D. Are they the same sensor really? Are they 'binned' for noise performance to determine what camera they go in (like most cpus are binned for speed for labelling)? Or are canon artificially crippling their noise performance somehow? (a 12 MP $400 1100D beating a $1500 18MP 7D wouldn't be good marketing, would it?) (But then, why wouldn't they cripple the 5D2 to be worse than the 1Ds3?).

No idea - but I'd speculate that Canon could easily 'dumb down' (or noise up, as the case may be) the firmware for the Digic processor, such that the RAW performance of the same sensor was 'tuned' for different cameras. Also, the AA filters may be of different strengths, etc. Having said that, DxOMark puts the ISO rankings at 784 and 854 for the 550D and 7D, respectively. Based on DxOMark's criteria, that difference is probably not significant.

They've just recombined the 1D/1Ds lines into the 1Dx for shutter speeds, wildlife/birders aren't too happy because now they have less pixels-on-target than the 1D4. Could they now introduce a 36MP 2/3/4D studio camera (to replace the 1Ds for high-MP), and a 5D3 at 18MP (1DX sensor) for see-in-the-dark (weddings/concerts/anything)? Then the birders can use the 2/3/4D and crop happily, as long as it inherits at least the old 1Ds 45pt AF and 6+fps...

They could release a 'pro' APS-C body (1-series styling and build) with improved sensor tech, and if they can convince those who 'miss' the 1D IV that the IQ of the new APS-C sensor surpasses the 1D IV, then they can sell that new camera for a bit under the price of the 1D IV, they'd probably have a winner on their hands.

No matter what they release, half the people on this board are going to complain.

+1

 

[Canon-F1]

Now, If you take a low-MP camera, and upscale the image to high-mp, you don't get nearly as much detail for printing billboards (sometimes I think everyone prints billboards at 300dpi on a daily basis given what I read here, yet noone gets paid enough for these billboards to hire a Phase One IQ 180).

LOL... so true!!! ;D

It´s like buying a sportscar in a country where you are only allowed to drive 80mph.
It´s more about "having it" then "using it".

Sad but true....


btw: sRAW is not demosaiced.

http://dougkerr.net/Pumpkin/articles/sRaw.pdf

Now, has “downsampling” actually taken place? No. We did not start
with a set of pixels at sensel resolution, and from it get a set of pixels
at half that resolution.
Nor did we start with a set of sensel outputs (R, G1, G2, B) at sensel
resolution, and get from it another set of corresponding values (R, G1,
G2, B) at half that resolution.
So we cannot speak of the “downsampling” of any consistent type of
data. Perhaps all we have is a very simple, direct way to develop a
pixel image from the suite of sensel data—one that requires no
“guessing” as to pixel colors.

Thus, we can see that perhaps what happens to the Raw data on its
way to the sRaw format does not, rigorously, involve either
demosaicing or downsampling.

(see page 8 of the PDF for explanation)

 

[jbwise01]

I like the direction Canon is headed with their DSLRs. I seriously can't believe all these posts about high ISO performance not being important. The truth is, High ISO performance is simply a higher priority for DSLRs going forward. As mentioned earlier, canon proved that they can easily expand MP count.

Benefits of High ISO development:
[list type=decimal]
[*]Better low light performance
[*]Expands capability of "Slower" lenses
[*]Allows higher shutter speeds to be used at narrower apertures
[*]Lens development can shift focus to increasing Sharpness overall without striving for wider apertures for low light performance
[*]Video performance is increased for all lighting conditions
[/list]

Let’s face it, this Move by Canon makes total sense. It’s not about increasing MP counts to appease the uninformed enthusiast; it’s about making the each pixel better. Once they have really pushed the limits of what we can expect out of each pixel, only then should they should move forward for high MP camera.

Its kinda like the car industry, they shouldn’t make cars with bigger engines with higher HP just because they can, they should be getting more out of the engines and HP they have to increase efficiency and performance of the car. The benefits increased efficiency far outweigh the benefits increased “power!”

 

[weixing]

Hi,
IMHO, you should compare using camera with different pixel size, but with the same technology and image processing unit. My suggestion will be using 5DII and 60D as both basically use the same technology and image processing unit... the only different is the pixel size.

Basically, took a similar image using under similar shooting condition and camera setting (should not be a problem since both are current Canon DSLR). Crop out 3479 x 2324 from the image shoot using 5DII... this should simulate a 8MP APS-C sensor (with pixel size 6.4um... a 350D using current sensor and image processing techology). Now resize the 7D 18MP to the same resolution (3479 x 2324) and compare the noise performance.

Just my $0.02.

Have a nice day.

 

[awinphoto]

Hi,
IMHO, you should compare using camera with different pixel size, but with the same technology and image processing unit. My suggestion will be using 5DII and 60D as both basically use the same technology and image processing unit... the only different is the pixel size.

Basically, took a similar image using under similar shooting condition and camera setting (should not be a problem since both are current Canon DSLR). Crop out 3479 x 2324 from the image shoot using 5DII... this should simulate a 8MP APS-C sensor (with pixel size 6.4um... a 350D using current sensor and image processing techology). Now resize the 7D 18MP to the same resolution (3479 x 2324) and compare the noise performance.

Just my $0.02.

Have a nice day.

Naw... then you're getting into the Full frame/crop sensor debate... crop vs crop, full frame vs full frame... 5d2 and D700 or if you want to stick with canon, 7d vs rebel xs? Heck even the 5d vs 5d2.. but then you got 2 different generations of cameras all together... but you get my point.

 

[weixing]

Hi,
IMHO, you should compare using camera with different pixel size, but with the same technology and image processing unit. My suggestion will be using 5DII and 60D as both basically use the same technology and image processing unit... the only different is the pixel size.

Basically, took a similar image using under similar shooting condition and camera setting (should not be a problem since both are current Canon DSLR). Crop out 3479 x 2324 from the image shoot using 5DII... this should simulate a 8MP APS-C sensor (with pixel size 6.4um... a 350D using current sensor and image processing techology). Now resize the 7D 18MP to the same resolution (3479 x 2324) and compare the noise performance.

Just my $0.02.

Have a nice day.

Naw... then you're getting into the Full frame/crop sensor debate... crop vs crop, full frame vs full frame... 5d2 and D700 or if you want to stick with canon, 7d vs rebel xs? Heck even the 5d vs 5d2.. but then you got 2 different generations of cameras all together... but you get my point.

Err... if you read carefully, it's not full frame vs crop... 5DII full frame sensor is 36mm x 24mm with 5616 x 3744 pixels, so a "single pixel" is around 6.4um. When you crop 3479 x 2324 pixels out of it, it's basically respresent a sensor size of around 22.3mm x 14.9mm (which is same as a 8MP APS-C sensor) using current technology. Please note that 350D had 3456 x 2304 pixels, so a "single pixel" is also around 6.4um.

The topics is about comparing noise performance of a shrinking High-MP vs Native Low-MP sensor, so in order to do a valid comparison, everything should be identical except sensor pixels count. So using 2 current DSLR from the same brand with different pixels size should be the best way to do it as the technology and image processing method between 5DII and 60D should be very identical except the sensor pixel size.

Have a nice day.

 

[Edwin Herdman]

It took me a while, but it hit me just now (having read Neuro's first post a long while ago) that something has obviously gone wrong if you're getting such different results from the same area of the sensor.

But I see now that's now what was done. After reading the first post's details of the setup (a few times) it strikes me as obviously bad methodology to compare a larger area from one sensor (1000 pixels square) with a smaller area (500 pixels square) in order to try to compare different sensor sites.

Or is something else meant by "cropped at 500x500" for the sRAW?

It seems to me that the only legitimate comparison would be between

1000x1000 and binned to 500x500 (keeping in mind the potential limitations neuro posted)
and
1000x1000 and not binned

I see this method was not chosen because of what I mentioned at the beginning (if you get wildly different results from binned / averaged / resampled pixels with respect to noise, something has gone terribly wrong), and it wouldn't tell us very much outside of the post processing context.

The other comparison doesn't make any sense to me, though.

 

[traveller]

Try this comparison between the NEX-5 and NEX-7 downsampled:

http://www.luminous-landscape.com/reviews/cameras/sony_nex_7_rolling_review.shtml#size-return

Probably a more meaningful test of a same-generation high resolution sensor downsampled versus a lower resolution sensor. That's not to criticise you dr croubie, I understand that you were trying to do your best with limited resources available.

 

[moreorless]

Try this comparison between the NEX-5 and NEX-7 downsampled:

http://www.luminous-landscape.com/reviews/cameras/sony_nex_7_rolling_review.shtml#size-return

Probably a more meaningful test of a same-generation high resolution sensor downsampled versus a lower resolution sensor. That's not to criticise you dr croubie, I understand that you were trying to do your best with limited resources available.

That seems to suggest Native > Downrezed to me, the NEX 7 is clearly inferior at ISO 800 and above.

 

[Edwin Herdman]

That seems to suggest Native > Downrezed to me, the NEX 7 is clearly inferior at ISO 800 and above.

Why would it suggest that? As the article states, when you downsample (in this case, I think you would call it "normalizing") a larger image it tends to help average out noise. I think it's just the case that the the higher MP camera came with a much larger tradeoff in the amount of noise in the signal, evidenced by the failure of downsampling to even it out. Some evidence for the "high MP cameras are too noisy" crowd, maybe.

 

[moreorless]

That seems to suggest Native > Downrezed to me, the NEX 7 is clearly inferior at ISO 800 and above.

Why would it suggest that? As the article states, when you downsample (in this case, I think you would call it "normalizing") a larger image it tends to help average out noise. I think it's just the case that the the higher MP camera came with a much larger tradeoff in the amount of noise in the signal, evidenced by the failure of downsampling to even it out. Some evidence for the "high MP cameras are too noisy" crowd, maybe.

One example isnt everything I agree but its one of the clearest I'v seen and thats with the NEX 7 sensor being 18 months newer than the NEX 5. You could argue its a case of Sony pushing too far too fast but then again 36-40 MP on FF would be pushing along way on from 21-24 we have now aswell.