5D4 Sensor Defect Discovered

[wockawocka]

Appreciate a lot if someone who tested the FW 1.03 could shortly say if there seems to be any changes in terms of the "phenomena".

Travelling, so my testing takes few days.

mnah mnah?

Exactly.

 

[PixelTrawler]

Within a laboratory where one is expected to draw firm conclusions based on data, mucking around in the near zero range of a signal is just plain wrong. Major ethical violation kind of wrong. People die kind of wrong.

+1E3

There may, in fact, be a problem in some 5D4 sensors; however, as the saying goes, the plural of "anecdote" is not "data." Seeing bands when pushing near-zero data is to be expected, and its absence would be a surprise.

Just a question, why would bands be expected? Theres no bands on the 5d3. I think the 5d3 is worse overall but the degradation noise is random with heavy pushes. If the streaks were random it would be much better.

I see FW 1.03 is now out also. It will be interesting to see if there are further improvements.

A few pages ago I posted a demo of a 16-bit blue gradient that I generated in Photoshop and then pushed the "exposure." It showed bands in the shadow area. I gave instructions to replicate it. For an explanation, see the Wikipedia article on posterization. In the bottom few bits of an image (dark areas) there is so little gradation that pushing the "exposure" will cause posterization, which shows as bands in some images.

Why not on 5D3: I can only speculate, but my guess would be the greater shadow noise on 5D3 turns the bottom few bits into a mush that does not display posterization as much when pushed.

Try this experiment: set up your 5D4 and 5D3 to capture a test image that should display the problem. Be sure to use a zoom lens rather than a prime. Find the right exposure to create the problem, then dial it in manually on both cameras. Now take 2 photos with each camera. Each set of 2 will have the same exposure, but the FL changed by just a few millimeters. The FL change should be enough to recognize, but not enough to make a dramatic difference to the overall image. Now push the exposures on all 4 images identically. Look at the two 5D4 images: are the bands in the same spot of the scene or the same spot of sensor? If it's the scene (i.e., they've moved relative to the sensor) then you have posterization. If they're in the same position in the frame then you might have a sensor-related phenomenon, though it may be a limitation rather than a defect. Now compare to your 5D3. Do you have the same image details at both ends of the DR? Same bands?

I'm not saying (and can't say) that there are no 5D4's with sensor problems; what I am saying is that pushing 5 stops will give you posterization in deep shadow areas.

I will try this tonight.

 

[Alex_M]

If you work with 16-bit file, pushing shadows should not cause posterisation. There is still enough gradations in 16-bit file even in the "bottom Bits". Plenty! As long as you not playing with the 8-bit jpg, it should not be an issue. You can, to some extent notice posterisation in shadow areas of the output file (jpg).
I have asked before and you have confirmed that you are seeing on-screen banding while you are working on 16-bit file in Adobe PS. Unless your screen resolution is set to 16-bit colours ( thousand of colours) it should not be the case.
Something is not right here with Adobe gradient rendering in my opinion.

Within a laboratory where one is expected to draw firm conclusions based on data, mucking around in the near zero range of a signal is just plain wrong. Major ethical violation kind of wrong. People die kind of wrong.

+1E3

There may, in fact, be a problem in some 5D4 sensors; however, as the saying goes, the plural of "anecdote" is not "data." Seeing bands when pushing near-zero data is to be expected, and its absence would be a surprise.

Just a question, why would bands be expected? Theres no bands on the 5d3. I think the 5d3 is worse overall but the degradation noise is random with heavy pushes. If the streaks were random it would be much better.

I see FW 1.03 is now out also. It will be interesting to see if there are further improvements.

A few pages ago I posted a demo of a 16-bit blue gradient that I generated in Photoshop and then pushed the "exposure." It showed bands in the shadow area. I gave instructions to replicate it. For an explanation, see the Wikipedia article on posterization. In the bottom few bits of an image (dark areas) there is so little gradation that pushing the "exposure" will cause posterization, which shows as bands in some images.

Why not on 5D3: I can only speculate, but my guess would be the greater shadow noise on 5D3 turns the bottom few bits into a mush that does not display posterization as much when pushed.

Try this experiment: set up your 5D4 and 5D3 to capture a test image that should display the problem. Be sure to use a zoom lens rather than a prime. Find the right exposure to create the problem, then dial it in manually on both cameras. Now take 2 photos with each camera. Each set of 2 will have the same exposure, but the FL changed by just a few millimeters. The FL change should be enough to recognize, but not enough to make a dramatic difference to the overall image. Now push the exposures on all 4 images identically. Look at the two 5D4 images: are the bands in the same spot of the scene or the same spot of sensor? If it's the scene (i.e., they've moved relative to the sensor) then you have posterization. If they're in the same position in the frame then you might have a sensor-related phenomenon, though it may be a limitation rather than a defect. Now compare to your 5D3. Do you have the same image details at both ends of the DR? Same bands?

I'm not saying (and can't say) that there are no 5D4's with sensor problems; what I am saying is that pushing 5 stops will give you posterization in deep shadow areas.

 

[Orangutan]

If you work with 16-bit file, pushing shadows should not cause posterisation. There is still enough gradations in 16-but file even in the "bottom Bits". Plenty! As long as you not playing with the 8-bit jpg, it should not be an issue.

If you have deep shadow you're effectively working with 8-bit depth. 8 bits gives you up to 255; max value of 16,383 per pixel for 14-bit capture. That's about 1.5%. Anything under 1.5% shadow is 8-bit depth. Ripe for posterization.

 

[tron]

If you work with 16-bit file, pushing shadows should not cause posterisation. There is still enough gradations in 16-but file even in the "bottom Bits". Plenty! As long as you not playing with the 8-bit jpg, it should not be an issue.

If you have deep shadow you're effectively working with 8-bit depth. 8 bits gives you up to 255; max value of 16,383 per pixel for 14-bit capture. That's about 1.5%. Anything under 1.5% shadow is 8-bit depth. Ripe for posterization.

This is interesting. Can you please elaborate? Also I believe/hope that the steps are not linear and that they would dedicate more bits for the lower end of the range but this is just speculation.

 

[Alex_M]

i am confused now:

I am talking about your Adobe PS 16-bit gradient that you have demonstrated.. It was created utilising full tonal range - from level 0 to level 255, white to black? if so, see attached image for what your gradient should look like on computer screen set to 24-bit colours.

If you work with 16-bit file, pushing shadows should not cause posterisation. There is still enough gradations in 16-but file even in the "bottom Bits". Plenty! As long as you not playing with the 8-bit jpg, it should not be an issue.

If you have deep shadow you're effectively working with 8-bit depth. 8 bits gives you up to 255; max value of 16,383 per pixel for 14-bit capture. That's about 1.5%. Anything under 1.5% shadow is 8-bit depth. Ripe for posterization.

 

[tron]

Hello,

I have upgraded my 5D4 to 1.03. There was no visible improvement in the test I did yesterday (same exposure settings, same setup). I now believe that our best hope resides in ... 5DMkV ( ;D or take your pick).

On a more serious note though we might have some hope in the form of 6DMkII (the way 6D was better than 5D3). Anyway Canon managers say that they use the best sensor available (they do not hold back sensor improvements).

EDIT:Initially I thought I saw an improvement. Then I realized that the composition was slightly different. When I fixed it the result was similar to previous day. Maybe I saw a slight improvement in the black jacket shots but they were not controlled in the sense I did not even use the same lens, setup composition etc. I just covered about 95% of the shot with the jacket and shot with -3 -3 2/3 and -4. The -3 photo had no banding. The -3 2/3 had very slight and the -4 had some. No a controlled test but a similar with the original. Anyway 5D3 with Magic Lantern destroys 5D4 in the -4 test (same composition, same shutter, speed, iso). No banding period!

I wonder what a Magic Lanterned 5D4 could achieve....

 

[Orangutan]

I am talking about your Adobe PS 16-bit gradient that you have demonstrated.. It was created utilising full tonal range - from level 0 to level 255, white to black?

No, it was a partial range, from dark blue to black. The point is to simulate the dark skies being lifted.

if so, see attached image for what your gradient should look like on computer screen set to 24-bit colours.

No, this doesn't apply. You've included too much range -- the point here is to focus on gradations of very dark blue.

Let's consider the bottom 8 bits in each of two 14-bit pixels. These pixels are <=1.5% of brightness. Assume that one of them is 255 (all ones in the bottom 8 bits) and the other is 127 (all ones in the bottom 7 bits). A 5-stop push is a factor of 2^5 = 32. So after the push, the brighter pixel is 8,160 and the dimmer is 4,064. The brightness gap has now changed from 128 (less than 1% of total DR, which is barely discernible) to 4096 (about 25% of DR). There's your posterization.

 

[scyrene]

I just read through the whole thread (I must be a masochist! Nah, it's fun). I conclude: a few 5D4's *may* have an issue. It's weird - reminds me a little of blooming from bright light sources on CCD sensors, but much more subtle. Tron seems to have been unlucky, if his 5D4's have an issue that makes them worse for his purposes than older bodies. That's a shame, but there's always gonna be some duds in a batch.

Tron aside, a very few others seems to have replicated this, but I agree with the overall impression that a lot of this is clutching at straws - the requirements for replicating this effect are so extreme (very dark exposure at low ISO, maybe a bright element in or near (one side?) of the frame, massive exposure lifting *and* shadow lifting, and according to some, a 'correctly calibrated' monitor) that it cannot be expected to show up in more than a few normal photos by a few photographers who specialise in nighttime landscapes (?). Given most of the complaints have come from a few vocal newbies, it's hard to judge. Intriguing that Neuro could process that shot of the building and get cleaner results with the same raw image and software. That suggests it's either not a sensor issue, or at least not wholly the camera, or at least that correct processing can address this.

Ultimately, there are too many variables to know what to think. Those 'true believers' have to accept that their asserions are not empirical and that even if they are unfortunate enough to have received defective cameras, this is such a marginal issue that it is at best of academic interest to most people, even those in the market for a 5D4 (which I possibly am).

The biggest impression I've got, however, is how amazing the 5D4 is for lifting shadows, compared to the 5D3 especially.

PS I doubt much will change with the next generation; if you're talking about conjuring data from almost zero signal, no technology can work miracles.

PPS One thing nobody seems to have provided data on is, whether the bands move, especially if light sources in or around the frame have moved. If it's completely fixed, you could maybe apply a dark frame technique akin to astrophotography. Either way, it would be interesting to know.

 

[scyrene]

Where's Dilbert on this subject? Seems strange without him!

Jack

I believe he was elected president.

 

[sherif]

am planing to buy this cam soon

 

[s2kdriver80]

I quickly perused through this thread. Unless I missed some of the posted photos, I haven't seen any with overexposure on the left and black on the right that shows obvious streaking other than the ones posted by the OP and mikethemaven.

I've seen a couple where the entire left was overexposed and the right was black, and the right side had banding all over the place, but that is to be expected from extreme pushes. Haven't seen photos like the ones from mikethemaven's examples where only part of the left frame was overexposed and where you can clearly see the streaking reflected on the right side, and the other test where the overexposed element is on the right and no streaking is shown on the left.

From what I understand, noise and general banding wasn't the issue; it was that streaking from an overexposed element on the left was reflected on the right side of the frame and not vice-versa. Maybe all 5D4s are like this. Whether or not Canon will address the problem is another matter.

 

[Bernd Stuht]

I see some green/auqamarine Colors in the Shadows, sometimes minimal as Stripes to the right.
BUT ! It seems to depend on DPP - when i exposure the RAW File in Camera the Effekt isn´t really noticable.

When i expose the RAWFile in DPP - the green/aquamarine Colors appear near really dark Area´s .
This False Colors are "movable" my changing the Shadows or Exposure.

The false Color ist visible in some normal - not pushed Pictures.

5D4 is really new, Firmware 1.0.3 -DPP 4.5.20.0

It would be possible to upload this RAWFile to my Server - if you would like to play with it - and see this Effekt.

Greetings Bernd