brainiac wrote:
He plainly states that the D3 is better at iso's above 1600, because he failed to uprez the D3 file to 21 Mpixels, and so misjudged his own crops.
Did he? It looks very much like the D3 pic is "uprezed" since it is much softer. Why would it be softer if both were 100% crops?
Where can one read about the comparsion?
Makten wrote:
Did he? It looks very much like the D3 pic is "uprezed" since it is much softer. Why would it be softer if both were 100% crops?
Where can one read about the comparsion?
I did the uprezzing in the crop above to illustrate that the original comparison on DPReview was a scientifically reliable way of handicapping the higher Mpixel camera. The original, with crops shown at different magnifications in order to handicap the 1Ds3, is here: http://www.dpreview.com/reviews/NikonD3/page32.asp
Uprezzing a file has little or no detrimental effect on real image data, as can be seen in a comparison of an uprezzed file and the original at 200%:
brainiac wrote:
I did the uprezzing in the crop above to illustrate that the original comparison on DPReview was a scientifically reliable way of handicapping the higher Mpixel camera. The original, with crops shown at different magnifications in order to handicap the 1Ds3, is here: http://www.dpreview.com/reviews/NikonD3/page32.asp
I see, but I don't think he had any intention to "handicap" the 1Ds. People reading dpreviews probably know enough to understand that a higher resolution camera with the same noise per pixel as a lower resolution camera, will have the lower total noise. Still, I can't see that the 1Ds has lower noise in your improved comparsion. I don't say it isn't better, but I can't see it because I don't know if it is noise or just the label that looks like that. Both cameras probably gives very low noise at ISO 3200, so a better scene for the comparsion is needed in my opinion.
Edit: And, JPG or RAW?
brainiac wrote:
Uprezzing a file has little or no detrimental effect on real image data, as can be seen in a comparison of an uprezzed file and the original at 200%:
While it is true that the crops look different, I would hesitate to say that the interpolated one lacks something that the original has.
Ehhh, what's the difference of "uprezzing" and doing whatever-it-takes to show it at 200%?? At 200%, the image IS "uprezzed". I don't get what you mean. You have just used two different types of interpolation.
Makten wrote:
People reading dpreviews probably know enough to understand that a higher resolution camera with the same noise per pixel as a lower resolution camera, will have the lower total noise.
So why does the reviewer say "There's no doubt that the D3 is the winner once you get above ISO 1600".
> Still, I can't see that the 1Ds has lower noise in your improved comparsion.
OK - I will try to explain what I see. The D3 has applied some chrominance noise reduction. That is why you see very limited local colour variation (colour detail) over the range of a few pixels. It's the familiar paint by numbers look. An example of this is that the chrominance NR thinks it has found a big green object and a big purple object in the tree, on roughly the 50 pixel wavelength. That's why there is a big green splodge in the top left of the tree, and a big pink splodge in the lower foliage. The 1Ds3 shows much more of the subtle colour variation there. The big splodges of colour are false, and due to an attempt to average chrominance noise in a way that identifies areas of uniform colour. You can duplicate this effect on Canon files in photoshop by applying the dust & scratches filter at increasing radii to the chrominance channels and not the luminance channel. The 1Ds3 crop is more accurate at retaining the variations of colour in the tree and elsewhere. In other words it needs LESS noise reduction to get an acceptable result, and therefore retains much more colour detail. Colour detail is a good thing, especially if you want a colour photo.
brainiac wrote:
Uprezzing a file has little or no detrimental effect on real image data, as can be seen in a comparison of an uprezzed file and the original at 200%:
While it is true that the crops look different, I would hesitate to say that the interpolated one lacks something that the original has.
Ehhh, what's the difference of "uprezzing" and doing whatever-it-takes to show it at 200%?? At 200%, the image IS "uprezzed". I don't get what you mean. You have just used two different types of interpolation.
No. The 200% non-interpolated file is the original pixels provided by the camera. The fact that in this case each one is represented by 4 monitor pixels should make no difference to the content of the picture, unless there is something very wrong with your monitor. From further away, the 200% crop I show should look identical to a 100% from closeup. This illustrates that interpolating upwards is not very destructive of image data, if at all, because the two crops look pretty much the same in terms of image detail.
brainiac wrote:
So why does the reviewer say "There's no doubt that the D3 is the winner once you get above ISO 1600".
Probably he has based that opinion on alot more pics than the example you showed, that really doesn't show anything.
OK - I will try to explain what I see. The D3 has applied some chrominance noise reduction.
How do you know that? And why wouldn't good NR be a relevant part of image quality? You can't really get the data raw, even when shooting RAW. There's always some processing going on out of your reach. Now, is this NR you're talking of done in PP, in RAW developement, or in the camera before even saving as RAW?
That is why you see very limited local colour variation (colour detail) over the range of a few pixels. It's the familiar paint by numbers look. An example of this is that the chrominance NR thinks it has found a big green object and a big purple object in the tree, on roughly the 50 pixel wavelength. That's why there is a big green splodge in the top left of the tree, and a big pink splodge in the lower foliage. The 1Ds3 shows much more of the subtle colour variation there. The big splodges of colour are false, and due to an attempt to average chrominance noise in a way that identifies areas of uniform colour....Show more →
Sorry, but this does not at all make my statement about the worthlessness of the test less true. I can't believe how anyone can rate those cameras with that example in mind. Though, it is kind of wierd that the reviewer chose those pics for making the difference clear. Or did you choose them?
You can duplicate this effect on Canon files in photoshop by applying the dust & scratches filter at increasing radii to the chrominance channels and not the luminance channel. The 1Ds3 crop is more accurate at retaining the variations of colour in the tree and elsewhere. In other words it needs LESS noise reduction to get an acceptable result, and therefore retains much more colour detail. Colour detail is a good thing, especially if you want a colour photo.
And how do you know what the Nikon file looks like without NR (If we suppose you are right about it)? I'd rather take less color detail than chroma noise, any day. So if Nikon filters it away, that's fine with me.
No. The 200% non-interpolated file is the original pixels provided by the camera. The fact that in this case each one is represented by 4 monitor pixels should make no difference to the content of the picture, unless there is something very wrong with your monitor. From further away, the 200% crop I show should look identical to a 100% from closeup. This illustrates that interpolating upwards is not very destructive of image data, if at all, because the two crops look pretty much the same in terms of image detail.
Alright, then I see what you mean. But I don't see the relevance.
Edit: And still: are those from JPG or RAW? That's like the most important question yet to be answered.
Makten wrote:
Probably he has based that opinion on alot more pics than the example you showed, that really doesn't show anything.
It is the only example he gave, and it shows that he is failing to assess images at equivalent output size. We have to assume that his opinion is based on the very same error he has made with his posted examples.
>> OK - I will try to explain what I see. The D3 has applied some chrominance noise reduction.
> How do you know that?
Because I can see it.
Luminance detail is of a far higher frequency than colour detail in the D3 crop. We know from previous Nikon bodies that Nikon has developed special in-camera noise reduction which blurs colour at high isos. You can see it. If you don't know what a mismatch between luminance detail and colour detail frequencies looks like, then I suggest you try applying chrominance dust & scratches to images in photoshop until you can spot the difference.
> And why wouldn't good NR be a relevant part of image quality?
I agree. That's why it makes sense to assess the two camera's outputs as they will be used. Appropriate use of noise reduction techniques or software makes sense in a comparison of very high iso shots, since it is often used in real work. It's important to show the non-NR'ed data too though. In this case we see a D3 file with chrominance NR, and a Canon without. The good thing about the D3 is that it can do that NR before saving a jpeg. Useful.
>> That is why you see very limited local colour variation (colour detail) over the range of a few pixels. It's the familiar paint by numbers look. An example of this is that the chrominance NR thinks it has found a big green object and a big purple object in the tree, on roughly the 50 pixel wavelength. That's why there is a big green splodge in the top left of the tree, and a big pink splodge in the lower foliage. The 1Ds3 shows much more of the subtle colour variation there. The big splodges of colour are false, and due to an attempt to average chrominance noise in a way that identifies areas of uniform colour.
> Sorry, but this does not at all make my statement about the worthlessness of the test less true. I can't believe how anyone can rate those cameras with that example in mind....Show more →
I am not defending the test, and I agree that the little information it does give us about the two cameras' outputs is flawed. I have tried to use the crops more appropriately to examine the real situation between these two cameras, i.e. what one would see in a print from each.
>> You can duplicate this effect on Canon files in photoshop by applying the dust & scratches filter at increasing radii to the chrominance channels and not the luminance channel. The 1Ds3 crop is more accurate at retaining the variations of colour in the tree and elsewhere. In other words it needs LESS noise reduction to get an acceptable result, and therefore retains much more colour detail. Colour detail is a good thing, especially if you want a colour photo.
> And how do you know what the Nikon file looks like without NR (If we suppose you are right about it)? I'd rather take less color detail than chroma noise, any day. So if Nikon filters it away, that's fine with me. ...Show more →
That's great - to each his own. However, the D3's in-built NR may be heavier than optimal in some cases, and too weak in others. Applying in post is a chore, but it does allow a very accurate result. I am not saying either approach is better or worse - it depends on the application, but I am saying that judging by these crops at iso 3200 the 1Ds3 is simply a better camera because its results are level with or better than the NR'ed D3 file for noise, while capturing a great deal more detail. Once the crops are shown at equivalent output sizes, I don't think many will prefer the D3 version.
>> No. The 200% non-interpolated file is the original pixels provided by the camera. The fact that in this case each one is represented by 4 monitor pixels should make no difference to the content of the picture, unless there is something very wrong with your monitor. From further away, the 200% crop I show should look identical to a 100% from closeup. This illustrates that interpolating upwards is not very destructive of image data, if at all, because the two crops look pretty much the same in terms of image detail.
> Alright, then I see what you mean. But I don't see the relevance. ...Show more →
The relevance is that it supports my claim that uprezzing the file with fewer Mpixels to the same size as the denser file is the best way to overcome the thoroughly misleading practice of comparing 100% crops. Once you have uprezzed the 12 Mpixel to 21 Mpixel, then 100% crops can be fairly compared, and will show you the real difference between the two cameras. In the case of these two cameras it will show that the stronger but smaller noise of the 1Ds3 is no worse than the weaker but larger scale noise of the D3. Incidentally it will also show that the 1Ds3 captures a whole lot more real information (detail).
OK, it seems we agree pretty well then, but what you've written above is something completely different from this, your second entry in this thread:
"My point is that Canon's best kept high iso secret is that the 1Ds3 is better than the D3 at high iso's."
I still don't believe it until I get to see raw results without NR, sharpening and such. What we can conclude is that the 1Ds probably is damn much better at high ISO's than many people claim. Maybe better than the D3. I don't know, and probably nor do you.
