Exmor vs DualISO

[raptor3x]

Note: This was a two particle article that I merged together so I may have missed some things and there may be sections where the writing doesn't flow properly.

I've already posted this on another website but thought people here might be intersted as well. I've noticed there's a feeling that MagicLantern's DualISO for the 5D3 is nothing more than a gimmick and that it's too flawed to actual be useful.I've been using it for over a year now and while there were some major tradeoffs to using it in the beginning, the MagicLantern guys (i.e. mostly A1ex) have made enormous steps towards making it a genuinely useful tool for high dynamic range shooting on Canon DSLRs.

When the A7R was first released I bought one because I wanted to see what it was like shooting on the same sensor as the D800E, but after a couple days I returned it because I just could not get over the terrible ergonomics and poor battery life of the camera. I did do some very rough comparisons between DualISO and the A7R at the time so I had some sense then of how they compared. Anway, fast forward to a couple weeks ago; like always happen someone was complaining that their 5D3 didn't give enough dynamic range, so I recommended DualISO and then a bunch of people started claiming that you lose half your resolution, get horrible aliasing and artifacts, it was completely unsuitable for printing, and still didn't get rid of the pattern noise or bring the DR anywhere close to an Exmor sensor camera. I wanted to create some kind of controlled experiment to compare a DualISO 5D3 with an Exmor sensor, as I hadn't seen anything like that online, so I bought an A7 and EF adapter to perform the test.

The first shot shows the two shots pictures with no post-processing exposure adjustments (Note: Please forgive the terrible picture, I was cleaning up my bike while I was doing this.) Exposure was set so that the sky was as exposed as possible without clipping and I turned off all the lights in my garage so that the interior was much darker. Focus is on the rectangular object in the lower left corner, which is a Risk box, and is ~5 stops underexposed. Both pictures were taken with a 70-200ii.

The second shot shows the result of a 5-stop push in LR to the entire exposure for each shot. The 5D3 shot was scaled to match the resolution of the A7 so there was no obvious difference due to resolution. Also, for these shots there's no luminance noise reduction applied but I did apply +25 of chrominance noise reduction; I'll show a shot with zero noise reduction in the full writeup.

Finally the meat of what I wanted to know, in a severely underexposed area of the frame where I pushed the shadows an ungodly amount how well does the 5D3 with DualISO compare.

The 5D3, left, shows a bit more noise and less detail but actually maintains more accurate colors than the A7 which seems to desaturate the shadows significantly. But compared to the difference between a vanilla 5D3 and the A7 and this becomes a pretty amazing result.

For fun, here's the vanilla 5D3 and the 1Ds3 for comparison.

Also, the X-T1 vs the A7.

 

[raptor3x]

I then took both cameras hiking to get some more real world type pictures as well as to show one of the weaknesses that I know still exists with DualISO. For all of the following pictures I metered so that the highlights were just clipping (based on the A7 zebras) and then dialed it down 1/3rd a stop. ISO, aperture, and shutter speed are identical for each scene, although the light was changing quite a bit so not all shots are exposed identically but I'd say they're all within 1/3rd stop. For the first shot I was using a 40mm STM on the 5D3 and a 16-35 F4 IS on the A7 (made the weight distribution fairly even and meant I didn't have to constantly swap lenses), while on the last shot I used the 16-35 at 16mm for both cameras. Both shots were done at ISO 100 while the secondary ISO for the DualISO shots was 1600 for the first shot and 3200 for the last shot were I wanted to show a specific type of artifacting that can occur.

The first is a tree that had a nice dark shadow down the middle from a branch hanging just right with a bright blue sky in the background.

Zooming in we can see that the shadow from the branch is almost completely black near the branch and then gradually lightens a bit.

First, comparing the vanilla 5D3 and A7 I pushed the exposure by 2 stops and set the shadows slider to +100 in Lightroom.

We can already see that the vanilla 5D3 is falling apart in the shadows with tons of color noise and mush for detail in the particularly dark regions while the A7 still looks fine. This was were the overall exposure looked fine but I decided to push two more stops just to see how the A7 did.

By this point the 5D3 shadows are unuseable garbage but the A7 is still looking respectable, very impressive and this is the primary reason why landscape photographers have been ditching Canon for Nikon over the past two years. Well that and the 14-24. And 36MP, but I digress.

So now that we've shown the A7 rules and vanilla 5D3 drools, how did magic lantern do? Again, we start off with +2 stop exposure and +100 shadows.

Whoa, huge difference compared to the vanilla 5D3. At this point I'll say the DualISO 5D3 and A7 are doing equally well, but what happens if we push another stop?

The two are still very close but I think the noise in the A7 has a slightly higher quality to it, still extremely close but I'll give it to the A7 at this point. How about one more stop.

Again, still close but I think the A7 is starting to pull away from the DualISO 5D3 now. The 5D3 seems a bit more exposed than the A7, probably from the changing light conditions, and the interpolation scheme seems to be showing some flaws at the edges of where the bark becomes washed out. Conveniently, Lightroom's CA removal tool works extremely well to get rid of these, but they can still show up if you push the files very hard. A win for the A7, but not by a large margin. Below, I'm showing the dualISO 5D3 next to the vanilla 5D3 to show what an insane difference dualISO makes for the 5D3.

 

[raptor3x]

The last thing I want to show is that there is another type of artifacting that can occur in situations where you push the secondary ISO very high, for this example 3200, and you have thin lines (i.e. small branches or wires) running almost parallel to the sensor rows. On the left hand side of the frame in the shots below you can see a tree with some bare branches sticking out.

If we zoom in, we can see some places where the tree branches seem to vanish and reappear. This can happen in a very specific scenario where you are :

• Using a very high secondary ISO. This usually only happens over ISO 3200, and MagicLantern will warn you about using these settings.
• You have a very thin object. Thin as in less than ~2 pixel in width.
• That thin object is at an angle of less than ~15 degrees to the horizontal row of pixels.

If you run into this scenario you will probably see this issue. Whether or not you'll see it in a print is another issue entirely. I made 13x19 prints of all the shots in shown here and I'd be surprised if anyone was able to tell the difference between the two without resorting to a loupe, even the issue with the branches is very hard to see in the prints. In addition, most scenes don't require the secondary ISO to be set any higher than ISO 800 as you really get diminishing returns for each additional stop of secondary ISO, but in cases where you absolutely need to push higher then be careful.

So who wins? Clearly the A7. It's simply the better tool for this kind of shooting. The workflow is way easier, you don't have to fiddle with installing MagicLantern, turning on DualISO, fiddling with the secondary ISO, or worrying about factors that might leading to artifacting. Plus it retains resolution better when pushed. That said, unlike with the vanilla 5D3 where the difference in image quality in these types of adverse situations is enormous, we're really nitpicking minor differences here. The fact that the guys from MagicLantern have extracted so much extra dynamic range from the Canon sensor is pretty amazing. Additionally, there's a yet to be released update for magic lantern that adds ~0.8 stops of highlight headroom for the 5D3 just using some of the pixel capacity that Canon decided to keep in reserve for whatever reason (i.e. when the 5D3 sensors clips to white, it's not actually clipped. There's at least another 0.8 stops available before it really clips.)

So is DualISO great for people who only shoot landscapes and run into DR issues all the time? I'd call it good but not great, if you're this kind of shooter you're crazy for not moving to Nikon or at least picking up an A7R. Is it great for people who normally shoot lower contrast scenes but get frustrated when they occasionally do run into a DR limited scenario? Definitely, this is pretty much the category that I fall into and DualISO has pretty much satisfied all of my DR issues (not that I wouldn't love a 5D4 with better low ISO performance).

As an aside, I like the A7 a bit more than I had expected but good lord are the ergonomics horrible. This is literally the least comfortable camera I've ever held. My wife shoots with a Fuji X-T1 and has much smaller hands so the grip doesn't bother her as much, but for me it's like Sony put an absolutely wonderful sensor inside of a torture device for my right hand. That said, the reason I went with Canon when I got my first DSLR was because the Canon models that I tried out fit my hand much better than any of the Nikons that I tried, so this issue probably won't be true for everyone.

 

[Cptn Rigo]

Thanks a lot for this info

Looks pretty interesting

I will try MagicLantern's DualISO

 

[Vossie]

Great article Raptor, with a nicely detailed comparison.

 

[sarangiman]

Nice work! Essentially, Magic Lantern is able to recover closer to the 'sensor DR' - by mitigating the effects of downstream read noise probably introduced en route to the off-imaging-chip ADC in Canon architectures.

The extra 0.8EV 'highlight headroom' (or sorts) is interesting. Makes me wonder if something similar is going on with the Nikon D810 & the 36MP sensor in there. DxO data suggest they're able to map a higher signal off the imaging chip to the same Raw value at ISO 64 -- essentially like extending the effective full-well capacity of pixels at this ISO.

 

[zim]

Very nice article, lots of effort to do that, most appreciated

Can these results be extrapolated to a 7D? I don't mean would a 7D be the same as these examples but would it have a similar ratio of improvement?

If so and this is all software based, makes me wonder when the (20.2MP guess) 7D2 makes an appearance and shows a similar improvement we will all be shouting praise be!!! and all 7D owners clamoring for an official software update ;D

Well if there is a marked improvement it will be easy enough to test how it was/wasn't done, just find a tree ;D

 

[pedro]

Intresting. Thanks for your effort. Therefore, ML would also provide to high ISO nightshots?

 

[zim]

Really surprised this topic hasn't generated more interest.

Doesn't this show in real world terms just how easy, without changing any sensor tech, it would be for Canon to improve one of the most complained about aspects of their current sensor design?

 

[East Wind Photography]

I've been using ML for quite a while now and use Dual ISO more than I really care to admit. However there is another round of tests you need to do. That is to use ISO 100 as your secondary or recovery ISO and 800 or 3200 as your primary.

Theoretically the results should be similar but there are reports of differences. I assume that it has to do with the number of bits recorded in the lighter part of the DR and less so in the shadows.

I started out using Dual-ISO with a high recovery ISO and discovered that I could not get the shutter speeds I needed for sports or even birds in flight. So I switched to shooting at ISO 800-1600 and recovering to ISO 100.

 

[East Wind Photography]

Really surprised this topic hasn't generated more interest.

Doesn't this show in real world terms just how easy, without changing any sensor tech, it would be for Canon to improve one of the most complained about aspects of their current sensor design?

What makes you think they haven't used this in the 1DX?

 

[East Wind Photography]

Very nice article, lots of effort to do that, most appreciated

Can these results be extrapolated to a 7D? I don't mean would a 7D be the same as these examples but would it have a similar ratio of improvement?

If so and this is all software based, makes me wonder when the (20.2MP guess) 7D2 makes an appearance and shows a similar improvement we will all be shouting praise be!!! and all 7D owners clamoring for an official software update ;D

Well if there is a marked improvement it will be easy enough to test how it was/wasn't done, just find a tree ;D

Dual-ISO works on every camera that ML supports (nightly builds only) including the 7D. However just note that you cannot push the ISO on the 7D much more than 800, even with Dual-ISO, or it really starts to fall apart.

I even use Dual ISO on my EOS-M!!!

 

[East Wind Photography]

Nice work! Essentially, Magic Lantern is able to recover closer to the 'sensor DR' - by mitigating the effects of downstream read noise probably introduced en route to the off-imaging-chip ADC in Canon architectures.

The extra 0.8EV 'highlight headroom' (or sorts) is interesting. Makes me wonder if something similar is going on with the Nikon D810 & the 36MP sensor in there. DxO data suggest they're able to map a higher signal off the imaging chip to the same Raw value at ISO 64 -- essentially like extending the effective full-well capacity of pixels at this ISO.

I guess the concept is that lower ISOs give more detail (bits per stop) in the highlights and higher ISO's more bits per stop in the shadows (sacrificing highlights for more shadow detail). Therefore it would seem correct that at ISO 64 you could extend the DR in the highlight end a bit more than Canon's ISO 100. Your quality of signal would determine what you get in the shadows.

Dual ISO just uses the best of both ISO's beneficial range and leaves the overlapped ranges untouched.

That brings up an interesting point though. Dual-ISO doesn't really yield lower resolution images. Only the highlights and shadows are half resolution and it's difficult to notice that based on how our eyes work. The rest of the image where the high and low ISO overlaps is essentially untouched and full resolution of your primary ISO. The wider the difference in ISO, the less overlap and therefore more potential to have your image de-res'd.

If you are shooting in daylight and want to recover shadows, use a high ISO of say 800 and a recovery ISO of 100. If you are shooting in low light and want to play around with HDR in that realm, you should increase your recovery ISO. I haven't tried low light HDR but it seems conceivable you could get better results than without Dual-ISO. Say using ISO 6400 and a recovery ISO of 800 or 1600. I'll have to play with that and see if there are any real benefits.

 

[dtaylor]

First off, to raptor3x - good article and great comparisons!

Doesn't this show in real world terms just how easy, without changing any sensor tech, it would be for Canon to improve one of the most complained about aspects of their current sensor design?

Complained about by who? Geeks arguing on forums? (No offense intended, I am a geek arguing on a forum.)

Even when you do have to push Canon shadows a little hard...shove the color NR slider over, all the way to the right if you have to, and give it some luminance NR. They're still not as good as Exmor shadows. But the difference is much smaller in a large print or stretched across a large monitor, and completely gone at average print and viewing sizes. The difference is never as large as it is while pixel peeping with minimal or no NR.

Just last night I was revisiting a Canon 7D landscape file with pushed shadows, not quite as hard as this test but hard enough. Pixel peeping on screen I can see the noise and it annoys me a bit. Printed to an Epson Ultra Premium Luster 17x22 sheet for one of my portfolio albums? I can't find any of that noise with my nose on the print.

Speaking of landscapes...high end professional landscape work is not produced by pushing Exmor shadows 4-5 stops. Landscape photographers bracket and HDR. Compare a HDR image to a heavily pushed image, even from Exmor, and the difference in tonality and fine detail will jump off the print at you. With AEB you can easily hand hold a 3 frame bracket.

All that said...I do find it puzzling that Canon went through the effort to make this possible in the sensor hardware but then never exploited it in the firmware. Are they afraid that it might be confusing to users, especially with the HTP mode option? Just add an Extended Dynamic Range (EDR) mode for RAW only and clearly state it's for pro users who are going to manipulate the tone curve in RAW.

It's dumb for Canon not to do this. But it is a much smaller issue, with far less impact on their bottom line, then any of us seem to realize.

 

[sarangiman]

What makes you think they haven't used this in the 1DX?

Because the 1Dx doesn't have much more DR than a 5DIII... and shows the typical pattern of increased total input-referred read noise with decreasing ISO.

 

[sarangiman]

Speaking of landscapes...high end professional landscape work is not produced by pushing Exmor shadows 4-5 stops. Landscape photographers bracket and HDR. Compare a HDR image to a heavily pushed image, even from Exmor, and the difference in tonality and fine detail will jump off the print at you. With AEB you can easily hand hold a 3 frame bracket.

Good point. And it's because of a subtle point. It's b/c even Exmor shadows are limited by shot noise - noise that comes from the inherent randomness of non-coherent light. This source of noise has increasing contributions with shorter exposures, even for an 'ideal' sensor (which the Exmor architectures are approaching).

You can think of it this way - if a landscape photographer won't use ISO 800, he/she similarly won't push a base ISO 100 camera's shadows 3 stops. Since it's essentially the same thing (ignoring quantization error).

All that said...I do find it puzzling that Canon went through the effort to make this possible in the sensor hardware but then never exploited it in the firmware. Are they afraid that it might be confusing to users, especially with the HTP mode option? Just add an Extended Dynamic Range (EDR) mode for RAW only and clearly state it's for pro users who are going to manipulate the tone curve in RAW.

Canon didn't go through the effort to make this possible in the sensor hardware. ML just found a way around the limitations of Canon's architecture, which pipes analog data off the imaging sensor to an off-sensor digitizing chip (ADC). This process introduces a lot of noise into the signal, which can be mitigated by pre-amplifying the image data (ISO amplification - raising the ISO) before sending it off the imaging chip. That's why Canon's architecture doesn't hurt its high ISO performance much, but severely limits its low ISO performance.

 

[verysimplejason]

I wonder how good will be a Canon 6D with dual ISO.

 

[raptor3x]

I wonder how good will be a Canon 6D with dual ISO.

Not a huge difference since the read noise difference between the 5D3 and 6D is neutralized and the 6D doesn't have a huge advantage at ISO 1600 over the 5D3.

 

[sarangiman]

Yeah, exactly what raptor3x said. You can probably safely assume a similar benefit across most Canon DSLRs.

Those guys at ML are magicians, I have to say. Personally, I think it's much easier to just go for a Nikon D810 if you care about base ISO DR, especially as the D810 offers yet another 2/3 EV or so DR over the D800/A7 cameras. But if you have other constraints that force you to stick with Canon, this ML trick seems very useful. As others mentioned, the resolution tradeoff doesn't seem as severe as one might expect it to be.

 

[sarangiman]

Intresting. Thanks for your effort. Therefore, ML would also provide to high ISO nightshots?

No, ML Dual ISO will not help with high ISO performance. Dual ISO simply helps to deal with the noise resulting from the architecture I mentioned earlier. The 'downstream' read noise this architecture is well known for - and what dual ISO helps to avoid - becomes pretty irrelevant at high ISOs. It's swamped by other sources of noise - namely shot/photon noise and 'upstream' read noise. Shot noise contributions are determined by how much light is collected, so roughly correlates with sensor size & QE. Nothing ML can do here. Upstream read noise for these Canon sensors is already quite good (especially for the 1Dx), and difficult to do anything about without actual hardware-level adjustments. So there's really nothing for ML to do here either.