speedmaster20d wrote:
Who is we? In my 10+ years of owning and shooting with Canon gear from 16mm to 1200mm I yet have to observe the significant variation internet experts on these forums talk about
From what I know so far, I really like this improvement that Canon has given us. I happen to enjoy images that look sharp, but of course hate artifacts that can creep in with normal sharpening. I am not saying that DLO will be perfect, but it does seem to provide a very optimum way to get from a RAW to a sharp image, but one that still can have most further normal processing done to it.
I also view DLO as in some ways equivilant to giving us more resolution in our cameras. The chase for the highest resolution we can get is mostly driven by the desire to accuratly get as much detail and sharpness as possible, and DLO might be roughly equal to 20% or 30% more mega pixels in the sensor. Time will tell as to if it really is that great, but it is clearly a step in the right direction.
And remember this is an improvement for many Canon cameras, not just the latest ones like the 5DIII. The quote Fred posted mentions it as working for cameras from the 30D onward, so a lot of people just got a retroactive upgrade that works on your old RAW files, and works with new RAWs from your current cameras.
mpmendenhall wrote:
Given a convolution with no zeros in the fourier transform of the kernel (e.g. a gaussian blur), in theory there is exactly one unique source image that will produce the convolved one. In the limit of a computer that can carry out calculations to infinite precision, the original image can be perfectly restored.
This theory breaks down in real life with the introduction of noise, both the noise from real camera images and quantization/rounding noise in computer calculations.
Yeah, spherical cows in vacuum have no problems with inverting convolutions... :-)
A relevant point: how do you distinguish noise in the image from fine detail? Say, I have a picture of a gravel road, that variation in pixel RGB values that I observe, is that noise, or is that fine detail of gravel? How can you tell without prior knowledge of what (parts of) the image is *supposed* to look like?
When an algorithm enhances small-scale variation in the sky we call it noise because we know the sky is supposed to be more or less even locally. When an algorithm enhances similar variation in a far-off forest we call it fine detail of foliage because we know that forests are not smooth. But that is knowledge from outside of the image, algorithms don't have it.
You're saying that the noise in the image spoils the perfect deconvolution -- isn't exactly the same as saying that the fine detail in the image spoils the perfect deconvolution?
Wow it's really quite stunning and matches the one other test I've seen so far.
The other test shot a much lower apertures and the difference still seemed to be large.
But man this will kill my workflow, I hate DPP. And I hate the NR in it, etc. But with such a huge difference I guess I'll have to use it for my ISO100-400 stuff with zero NR and basically everything turned off other than DLO and then export to photoshop. I don't like the exposure controls and so as much but for that improvement I guess I'll have to live with DPP and exporting and all that hassle.
I'm downloading it now. If my results are the same, and for lower apertures too, and the difference will be that great vs best effort with ACR, which remains to be seem, then wow, that seems almost like the difference between 22MP and maybe 30MP!
skibum5 wrote:
Wow it's really quite stunning and matches the one other test I've seen so far.
The other test shot a much lower apertures and the difference still seemed to be large.
But man this will kill my workflow, I hate DPP. And I hate the NR in it, etc. But with such a huge difference I guess I'll have to use it for my ISO100-400 stuff with zero NR and basically everything turned off other than DLO and then export to photoshop. I don't like the exposure controls and so as much but for that improvement I guess I'll have to live with DPP and exporting and all that hassle.
I'm downloading it now. If my results are the same, and for lower apertures too, and the difference will be that great vs best effort with ACR, which remains to be seem, then wow, that seems almost like the difference between 22MP and maybe 30MP!
wickerprints wrote:
It seems that according to the pre-release instruction manual, at the present time, only the following Canon lenses are supported (though this may change before the official release of the next version of DPP, and it does not take into account future profiles Canon might decide to create):
EF 14/2.8L II USM
EF 24/1.4L II USM
EF 35/1.4L USM
EF 50/1.4 USM
EF 50/1.2L USM
EF 85/1.2L II USM
EF 300/2.8L IS II USM
EF 400/2.8L IS II USM
EF 500/4L IS II USM
EF 600/4L IS II USM
EF 16-35/2.8L USM, II USM (both versions)
EF 17-40/4L USM
EF 24-70/2.8L USM, II USM (both versions)
EF 24-105/4L IS USM
EF 28-300/3.5-5.6L IS USM
EF 70-200/2.8L IS USM, II USM (both versions)
EF 70-200/4L USM, IS USM (either with/without IS)
EF 70-300/4-5.6 IS USM [not the L?]
EF 100-400/4.5-5.6L IS USM
EF-S 10-22/3.5-4.5 USM
EF-S 15-85/3.5-5.6 IS
EF-S 17-55/2.8 IS USM
EF-S 17-85/4-5.6 IS USM
EF-S 18-200/3.5-5.6 IS
EF-S 18-135/3.5-5.6 IS
I think the range of supported lenses suggests that Canon chose designs that (1) they felt would benefit from this algorithm, and/or (2) are especially popular lenses, and/or (3) tend to be used by photographers who wish to extract the maximum performance out of the system.
I am, however, somewhat surprised that the macro lenses were not listed. The EF 100/2.8L macro IS, EF 180/3.5L, and MP-E 65/2.8 1-5x macro are frequently used stopped down for maximum DOF. And to some extent, I could also imagine DLO being useful for the TS-E lenses, so it's a shame we don't see those in the list, either. It's entirely conceivable that we might see more Canon lenses supported in the future, though.
As for third-party lenses, I think it would be impractical for Canon to spend the resources to measure a competitor's lens and create profiles. It may not even be feasible, since for all we know Canon could be using their own ray traced simulations of the spot diagrams of their lenses to compute the profiles.
As for me, I have 3 lenses in that list--4 if you count ones I've since parted with but from which I still have images. ...Show more →
hmm bizarre to leave off the macros
and the old 70-300 IS but not the L
and no love for the old 300 2.8 IS with or without TC
skibum5 wrote:
Wow it's really quite stunning and matches the one other test I've seen so far.
The other test shot a much lower apertures and the difference still seemed to be large.
But man this will kill my workflow, I hate DPP. And I hate the NR in it, etc. But with such a huge difference I guess I'll have to use it for my ISO100-400 stuff with zero NR and basically everything turned off other than DLO and then export to photoshop. I don't like the exposure controls and so as much but for that improvement I guess I'll have to live with DPP and exporting and all that hassle.
I'm downloading it now. If my results are the same, and for lower apertures too, and the difference will be that great vs best effort with ACR, which remains to be seem, then wow, that seems almost like the difference between 22MP and maybe 30MP!
wickerprints wrote:
Some things I read from the manual:
1. No lens+extender combinations are currently supported, which is a shame considering that it's when we use the extenders that the performance of the super-telephoto primes suffer enough that it might be worth trying to correct! I've never taken a shot with my 300/2.8L IS that made me think, "oh, if only the lens aberrations/diffraction were not so prominent."
yeah it is confusing that they leave out lens+extender and macros where it could be a of great benefit plus the crazy $$ 70-300L gets no love? maybe they feel they didn't charge enough for it
and i hope they don't try the trick that the old super-tele are obsolete so they will never get support, love to see 300 2.8 IS + 2x TC III (and 14x TC III and bare lens). IMO that would be nuts since few are going to rush out for the $7000-$15,000 lenses and they'd be giving back the reach advantage, in full, to Nikon with their D800.
speedmaster20d wrote:
Canon are doing actual measurements to calibrate each and every lens as opposed to using their simulated design data.
Oh, and by the way, doesn't it mean that anyone with enough money to buy a bunch of Canon lenses and some sophisticated-but-not-exotic equipment can do exactly the same thing?
Fred Miranda wrote:
It would be really helpful if we could batch our images through DLO alone and have a resulting RAW file which could then be edited in LR or ACR.
but the f/16 samples from the OP really did look like a 5D3 vs D800 difference in MP, it will be interesting to see how extreme things are at f/2.8-f/7.1
KaaX wrote:
Yeah, spherical cows in vacuum have no problems with inverting convolutions... :-)
A relevant point: how do you distinguish noise in the image from fine detail? Say, I have a picture of a gravel road, that variation in pixel RGB values that I observe, is that noise, or is that fine detail of gravel? How can you tell without prior knowledge of what (parts of) the image is *supposed* to look like?
When an algorithm enhances small-scale variation in the sky we call it noise because we know the sky is supposed to be more or less even locally. When an algorithm enhances similar variation in a far-off forest we call it fine detail of foliage because we know that forests are not smooth. But that is knowledge from outside of the image, algorithms don't have it.
You're saying that the noise in the image spoils the perfect deconvolution -- isn't exactly the same as saying that the fine detail in the image spoils the perfect deconvolution?
Note that the smooth sky is smooth and the details are detailed. The algorithm is dumb as a sack of rocks plus FFTW3, yet it "knows" how to handle each area without outside knowledge. This is because, when noise is sufficiently low, deconvolution can be done nearly perfectly with a rather crude approach --- the necessary information to distinguish detail from smoothness is still present in the blurred image, until masked by noise.
My reason for identifying the role of noise as the key limiting factor is to be informative about where one of the key difficulties in implementing such procedures "in real life" lies. Even the smallest amount of ISO 100 camera noise will go from "who cares" to "a big deal" as this type of correction becomes more popular, and is pushed to harder limits by high-resolution camera sensors.
Fred Miranda wrote:
It would be really helpful if we could batch our images through DLO alone and have a resulting RAW file which could then be edited in LR or ACR.
I am not sure about batching, Fred. The file size will increase by 3X. I would do a pick and choose method only. The adjustment will be incorporated in the new RAW file and LR or any other can pick it up from there.
mpmendenhall wrote
My reason for identifying the role of noise as the key limiting factor is to be informative about where one of the key difficulties in implementing such procedures "in real life" lies. Even the smallest amount of ISO 100 camera noise will go from "who cares" to "a big deal" as this type of correction becomes more popular, and is pushed to harder limits by high-resolution camera sensors.
Thanks for pointing this out, it does make perfect sense though.
mpmendenhall wrote:
This is because, when noise is sufficiently low, deconvolution can be done nearly perfectly with a rather crude approach --- the necessary information to distinguish detail from smoothness is still present in the blurred image, until masked by noise.
I don't think you responded to my point which is that deconvolution algorithms are not capable of distinguishing detail from noise.
"Masked by noise" is exactly the same thing as "masked by fine detail". If deconvolution breaks down in the presence of noise, it also breaks down in the presence of fine detail.
I think the issue is the matter of scale, or, rather, of the relationship between three magnitudes -- the magnitude of the convolution (the "strength"/diameter of the blur), the size of the features we're interested in (be they noise or fine detail) and the size of the underlying discrete pixel.
If, for example, features are large and the magnitude of the convolution is small, deconvolution proceeds well (e.g. make a large sharp edge that got blurred sharp again). However if the convolution is large (lots of blurring) and the features are tiny, well, let's say it doesn't work all that well.
KaaX wrote:
I don't think you responded to my point which is that deconvolution algorithms are not capable of distinguishing detail from noise.
"Masked by noise" is exactly the same thing as "masked by fine detail". If deconvolution breaks down in the presence of noise, it also breaks down in the presence of fine detail.
I think the issue is the matter of scale, or, rather, of the relationship between three magnitudes -- the magnitude of the convolution (the "strength"/diameter of the blur), the size of the features we're interested in (be they noise or fine detail) and the size of the underlying discrete pixel.
If, for example, features are large and the magnitude of the convolution is small, deconvolution proceeds well (e.g. make a large sharp edge that got blurred sharp again). However if the convolution is large (lots of blurring) and the features are tiny, well, let's say it doesn't work all that well. ...Show more →
I was hoping that my images would count for a thousand words of response.
Earlier in the thread, you asked for someone to prove that a deconvolution could work "perfectly" under near ideal conditions (synthetic image with no added noise, blur kernel perfectly known). That's exactly what I've done, with the images in my posts as demonstration. I've shown, under ideal conditions, the ability to de-convolve a significantly blurred image, using a completely "dumb" algorithm that somehow still manages to distinguish smooth sky from fine textures. I've also tried to show how this ideal case breaks down with the addition of small amounts of noise (which does, indeed, mask real detail due to texture --- but I've demonstrated that without the added noise, texture/smooth is fully recoverable). What exactly do you want demonstrated that I haven't already shown through the images posted?
mpmendenhall wrote:
What exactly do you want demonstrated that I haven't already shown through the images posted?
I want you to recover the texture (= fine detail, NOT the difference between smooth areas and textured areas) from a significantly blurred image. The claim was that Gaussian-blur convolution is *perfectly* invertible.This means that given a synthetic image with no noise you should be able to recover a pixel-perfect copy from a highly blurred version (ignoring the edges) -- right?
KaaX wrote:
I want you to recover the texture (= fine detail, NOT the difference between smooth areas and textured areas) from a significantly blurred image. The claim was that Gaussian-blur convolution is *perfectly* invertible.This means that given a synthetic image with no noise you should be able to recover a pixel-perfect copy from a highly blurred version (ignoring the edges) -- right?
And that is exactly what I've already posted. Compare my deconvolved images to the original: all the fine detail/texture is (nearly) perfectly reconstructed. Look at my blurred picture: I'd call that sufficiently blurred to obliterate visible traces of the detail/texture (look, e.g., at the tree bark or the red tile roof upper left). What's missing?
