[Rondec]

Well, on digital a pixel is a pixel. Could be of different size on the sensor but once sampled to data it does not have a size anymore just a value. So you just have data. It could be 16mpx from aps-c sensor or 24 from FF sensor or something else but it does not matter anymore since it only measures the noise based on the pixel values. As the output pixels are the same size if you use the same output media for both images then really more mpx = bigger source image size.
Then when you know your what is your raw noise + your resolution and this gives you some idea wether you need to denoise and how much without losing too much details, how much you can scale and how much potentially you can crop to get to your target output media resolution of 8mpix while keeping sufficient image quality. If you just look at a normalized value then these things are harder to tell as differences are averaged out when you scale both images down. Just to take it to the extreme if you scale the noisiest possible 4mpix image and the cleanest possible say 24 mpix image to a thumbnail of 16x16 px probably you can't measure any difference anymore off the thumbnais. Then try to scale the same to 8mpix ...

Anyway, It is pointess to compare 2 different mpx sensors just for noise without considering their resolution as well and normalizing the images to a third size does not tell the whole story.
For same mpix sensors regardless of their size there's no need to normalize since they are already "normalized".

People compare output of sensors all of the time. "My K5 produces lots better images than my K7 did," or whatever. This is done without specific measurement, but more with a feel to how much you can push the RAW files you get from a given sensor at a given iso and then what size you can print that file at.

As to the fact that DXO Mark picked 8 megapixels as their normalization size, when they started their web site, 6 megapixel sensors were still pretty common. At this point, obviously most cameras have higher numbers of megapixels, but to adjust the normalization size, say, to 16 megapixels, would make it impossible to backwards compare cameras tested with the new normalization standard to those tested with the old one.

In the end, it is all about the print or the viewed image. The assumption has to be made that if you were taking a given image with whatever camera, your end size would be the same. Clearly in that situation, some sensors will perform better than others.

[derekkite]

They normalize to a 300dpi 8*12 image.

It is an answer to what happens when you take an image from a 24mp apsc and a 32mp ff, output to the same size and resolution.

That is all. Whether it is helpful or representative of anything is an open question. Typically better noise characteristics and higher resolution give better results in their calculation.

I think it is helpful but it doesn't reflect some of the subtleties of different sensors, but generally is representative. Most of us don't do 18% grey images.

[jsherman999]

As to the fact that DXO Mark picked 8 megapixels as their normalization size, when they started their web site, 6 megapixel sensors were still pretty common. At this point, obviously most cameras have higher numbers of megapixels, but to adjust the normalization size, say, to 16 megapixels, would make it impossible to backwards compare cameras tested with the new normalization standard to those tested with the old one.

The point they make in their narrative description is that it wouldn't matter what arbitrary MP target you normalize to - the plots would still be the same, only shifted by a constant.

Again - this means that normalization to 4, 12, 24, 54MP, whatever - the curves wouldn't change relative behaviour and Camera A will never 'catch' camera B if it didn't at 8MP.


.

[jsherman999]

No, this is a common misconception. In a same-gen comparison, a larger sensor with greater pixel density than the smaller will still show less noise. It's about sensor area there, along with the mandated larger physical aperture at the same exposure for that FOV. Really, really, it is I can provide some external references if you wish for further reading ....

Here's another good intro: the effect of pixel and sensor sizes on noise

...The total amount of light that goes to make up your image is the most important factor in determining image quality....
The assertion that a smaller sensor sees less light and is, therefore, noisier (since it receives less light, and hence has a worse signal-to-noise ratio because of shot noise), is contentious. This is understandable, as it means thinking in terms of the whole image, not the light intensity per unit area that underpins the f-number/ISO system of exposure (that is specifically designed to work independently of sensor size).

However, while it may seem heretical to anyone strongly attached to the conventional exposure model, it is, nonetheless, demonstrably true.*

And what follows in that article is an *excellent* mouse-over demonstration of the effect of normalization of the image, both normalization of MP and greater sensor area. It will show you in image exactly what the 'print' tab shows in graph - for comparing sensors of different sizes through image size normalization - which is one of the things the print tab provides.

...Now we see the effect of all the extra light captured - the outer cone in the diagram at the top of the page. The APS-C camera and APS-C crop still look the same but the extra sensor area (all exposed to the same light intensity) allows the full frame image look better overall.

Note that the full frame sensor performs better than the APS-C sensor, even though its pixels are not bigger. If you were to take full frame camera with the same pixel count as the APS-C one (and so have bigger pixels), the result would still be very similar indeed.

.

---

Last edited by jsherman999; 03-18-2016 at 06:18 PM.

[npc]

The point they make in their narrative description is that it wouldn't matter what arbitrary MP target you normalize to - the plots would still be the same, only shifted by a constant.

Again - this means that normalization to 4, 12, 24, 54MP, whatever - the curves wouldn't change relative behaviour and Camera A will never 'catch' camera B if it didn't at 8MP.


.

Image is made of discrete pixels and the pixels have discrete and finite number of values they can take per channel. So in theory, if you scale the images enough the difference gets small enough and pixels will eventually just clip to the same value and the algorithm will get the same measurement from both images. At this point camera A "catches" with camera B.
How much you need to scale until this happens is another story but the smallest size you can possibly scale an image is to 1 px - my uneducated guess is all channels will be clipped to the same values on both images at that point

[jsherman999]

Image is made of discrete pixels and the pixels have discrete and finite number of values they can take per channel. So in theory, if you scale the images enough the difference gets small enough and pixels will eventually just clip to the same value and the algorithm will get the same measurement from both images. At this point camera A "catches" with camera B.
How much you need to scale until this happens is another story but the smallest size you can possibly scale an image is to 1 px - my uneducated guess is all channels will be clipped to the same values on both images at that point

Correct for downsampling past a certain point. I was kinda continuing on the line of your thought of scaling up - suggestion being that the relative performance of the cameras might change at some arbitrary upscale point greater than dxos 8mp. DXO shows their formula in the narrative and describes why that wouldn't happen.

I think your overall point about different processing has validity, I'm just not sure what a DXO can do about it or how they could represent it beyond the combination of screen/print tabs they already have.

[derekkite]

Remember what is scaled is noise. We know intuitively that in a situation such as a sky or a mountainside covered with trees, there is either lots of surrounding data to smooth the noise or lots of irregular detail that hides it.

But if you have a shot of very small print on a page or an eye of an animal or person, those lost pixels can't be extrapolated away. Is the s a g? Or the B an R. Or what is the eye color? The more pixels available the more likely these details will be preserved. A sensor with 20% more pixels but 5% more noise may in some circumstances give better end results, all else being equal.

That is the dynamic the print tab is trying to represent. I wouldn't vouch for it as a hard and fast rule, in fact it is based upon a photo of 18% grey. That doesn't represent my photos.

---------- Post added 03-19-16 at 08:12 PM ----------

In audio equipment the effect of SNR depends on what and how you listen. A very quiet music piece with subtle playing needs a higher SNR than loud rock and roll. The composition of our images and how much we crop and/or magnify will change the effect of noise and resolution.

[Ian Stuart Forsyth]

Remember what is scaled is noise. We know intuitively that in a situation such as a sky or a mountainside covered with trees, there is either lots of surrounding data to smooth the noise or lots of irregular detail that hides it.

But if you have a shot of very small print on a page or an eye of an animal or person, those lost pixels can't be extrapolated away. Is the s a g? Or the B an R. Or what is the eye color? The more pixels available the more likely these details will be preserved. A sensor with 20% more pixels but 5% more noise may in some circumstances give better end results, all else being equal.

That is the dynamic the print tab is trying to represent. I wouldn't vouch for it as a hard and fast rule, in fact it is based upon a photo of 18% grey. That doesn't represent my photos.

---------- Post added 03-19-16 at 08:12 PM ----------

In audio equipment the effect of SNR depends on what and how you listen. A very quiet music piece with subtle playing needs a higher SNR than loud rock and roll. The composition of our images and how much we crop and/or magnify will change the effect of noise and resolution.

Noise reduction loves resolution

Sounds about right (some noise reduced samples): Open Talk Forum: Digital Photography Review
a very good read with regard to captured resolution and how NR can clean up your images

[Rondec]

Correct for downsampling past a certain point. I was kinda continuing on the line of your thought of scaling up - suggestion being that the relative performance of the cameras might change at some arbitrary upscale point greater than dxos 8mp. DXO shows their formula in the narrative and describes why that wouldn't happen.

I think your overall point about different processing has validity, I'm just not sure what a DXO can do about it or how they could represent it beyond the combination of screen/print tabs they already have.

The thing is that mathematically, an APS-C camera will never "catch" a full frame camera -- regardless of how much you down sample the images. But we know that there are print sizes below which it would be hard to tell the difference based on resolution alone. Print a bunch of photos at 5 by 7 from a D810 and a K5 II and other than depth of field, it is probable you can't tell the difference. Print them at 16 by 20 and it is much more likely the increased detail in the D810 photo will show up on close inspection.

[jsherman999]

The thing is that mathematically, an APS-C camera will never "catch" a full frame camera -- regardless of how much you down sample the images. But we know that there are print sizes below which it would be hard to tell the difference based on resolution alone. .

Think that's what we were saying there re downsampling.

Re upsampling, it can't 'catch' it either, but what would happen there is some larger point they would both suck so bad you wouldn't care which one was minimally better. You'd be ready for more MP

[derekkite]

Think that's what we were saying there re downsampling.

Re upsampling, it can't 'catch' it either, but what would happen there is some larger point they would both suck so bad you wouldn't care which one was minimally better. You'd be ready for more MP

This is the conclusion I've reached. A friend who shoots Nikon has gone through this a few times considering the D810 vs the D7200 he shoots now. Mostly long lenses. We were talking the other day, on one of our expeditions. This time landscape and spring flowers. Essentially for a long cropped shot the same results would come if the crop was 20% larger in % of full frame sensor area on the K1. But most of the time those shots are trash anyways.

And that was the point. I know I'll never do anything except keep as a record the flock of 30 Western grebes way out in the middle of the lake. All I want is enough detail to identify and count them. But the ones I print are almost always close in, and I've missed more very close in shots where either the AF couldn't keep up or the subject was larger than the frame than ones where nothing would get the shot in any case. And the family of grizzlies in the morning light is possible with the low light performance and wider field of view. Once the light came up you couldn't miss no matter what you shoot with.

The other attributes of the sensor and body will end up being positive as well. I know he would have a D810 if it was under $2k. I'll probably end up with the K1 as well, and keep the K3.

[Winder]

I took images form my old 12MP Canon 5D and my 12MP Olympus (apples to apples). I processed them both on my overclocked Commodore 64. All images were checked for color accuracy on my calibrated IBM 8513 VGA monitor. On screen both images look identical so I decided to put them to a print test. I put in a fresh ribbon in my Okidata 320 Turbo printer and made 8x10 prints of each. Then, in order to get a completely unbiased opinion I call my wife in to be the judge and tell me which one she thought was APS-C and which one was FF. She looked at my like I was crazy. I don't think she wants to get involved is such a socially sensitive issue in an election year. I forgot to label the prints, so I don't actually know which is which now, but even the carbon copies would look good at 50 feet. Noise is virtually non-existent even at 1:1.

[jsherman999]

I can't wait to upgrade to a C64, I'm still using my Vic-20. 36mp is a b***h, it takes me 400 cassette tapes to store one image.

[Tan68]

... Noise is virtually non-existent ...

B/S

The look was your wife trying to figure out who had been using a jackhammer to make prints...