Magic Lantern - Dual ISO readout

This is pretty slick - I am going to try this over the next week or so and see how it works. THe forum write up looks pretty promising. The PDF documentation that is downloadable from the forum starts like this:

Pretty intriguing!

Here is the link to the ML Forum post:

http://www.magiclanter​n.fm …ic=7139.msg5968​7#msg59687

In essence, ML permits you to set up your exposure so that, for a single click of the shutter release, half of the sensor data is acquired at ISO 1 and the other half acquired at ISO 2. So, for example, you could use ISO 100 and 800 and shoot a scene that will read off half of the sensor at ISO 100 and the other half at ISO 800 (alternate sensor rows - actually alternating every two sensor rows to get the complete RGGB sensel block). This means you are capturing a single image, including motion blur and moving objects with two exposures simultaneously. The tradeoff is half vertical resolution, as the above quote indicates, as well as aliasing in the highlights and shadows, where the skipping of lines is more evident.

I look forward to trying this.

The workflow required to convert your raw image to something useable is provided by additional applications or adaptation of existing open applications (like EXIFTool and dcraw).

Previous extension of DR with ML (especially for video) used alternating ISOs but per frame - that is, capture frame 1 at ISO 1 and frame 2 at ISO 2 and combine these two frames into a single frame and repeat. This leads to some complex frame blending issues that have been pursued and solved in novel ways, thanks to John Hable (GingerHDR). A1ex's dual ISO readout is an equally clever approach that promises even better coping with motion with extended DR.

A1ex never ceases to amaze.

Enjoy!

kirk

I've been using this for a couple of months. It's impressive and the software to convert to a .dng has improved significantly to the point where it can be really hard to see any resolution differences. I need to test a bit more for high-res landscapes, but even for that it's looking pretty good.

A recent test. Full image processed as similarly as I could (i.e. I did try to bring out the shadows on both) and 100% crops (unsharpened) with a +1 EC to see the dark details. I don't think I need to tell you which is which.

Looking at this I find it hard to see the loss of resolution and obviously with some sharpening - well one could actually sharpen the DualISO image in the shadows. Just goes to show what the Canon sensors are really capable of.

HOSTED PHOTO
please log in to view hosted photos in full size.

HOSTED PHOTO
please log in to view hosted photos in full size.

In essence, this technique is an in-camera version of the Zero Noise concept we discuss here every once in a while. If you read A1ex's white paper:

http://acoutts.com/a1e​x/dual_iso.pdf

This is sufficient to extend dynamic range for about 95% of the shots I take, making time-consuming HDR acquisition even less necessary, with better results. I know that when I was emailing with Guillermo as I was experimenting with Zero Noise, his goal was ultimately to output a DNG instead of a TIFF. Essentially, this is what you get from the ML dual-iso image plus the conversion in the cr2HDR application.

I've been "learning" how to use it, and I am still experimenting with getting the best initial metering for the dual ISO capture.

Cool stuff, thanks for sharing!

kirk

If I remember reading correctly, 600D also supports this?

kirkt wrote in post #16343323
I've been "learning" how to use it, and I am still experimenting with getting the best initial metering for the dual ISO capture.

kirk

Originally, the conversion software didn't seem to work very well if the camera ISO was higher than the DualISO, but the latest version seems to do just about the same. The only difference is the camera calibration I have to use in ACR.

So I usually am shooting for the shadows and I set the DualISO to 100 to preserve highlights as much as possible. This also works with flash. If the subject is in the shadows and you want to set the camera ISO to 100 and use 800 or 1600 in the ML menu, then what I have found to work is just set EC to -3EV. If using a flash I just set the flash EC to -3EV as well +/- whatever I'd normally set. Normally now I just have the ML menu ISO set to 100 though.

For landscape and the above shot example, where I absolutely do not want to clip highlights, then I just use the raw histogram, set ISO100 and make sure the highlights are not clipped. Then set the DualISO to 800 for my 7D or 1600 for the 5DII.

I did have to create camera profiles for both cameras for DualISO and also different ones for the 100/800 and 800/100 combinations.

If you end up doing something different, I'd certainly be interested.

The raw histogram is really a huge help. I have pretty much started to home in on a similar approach to yours. I may have time today to take some test shots and see how they come out. I also typically have been using 100-800 or 100-1600 on my 5DII. I have been experimenting with various approaches of converting the processed DNG into an RGB image for editing. Depending upon the DR of the scene, the post-processing is very similar, for obvious reasons, to that I use for Zero Noise images. That is, the processed DNG is scaled so that the image resembles the shot taken at the lowest ISO (highlights preserved, everything else very dark).

As I start to get a better grasp on the workflow and the challenges of using the dual iso images, I'll post back here too.

kirk

Here is an example of an image I shot this weekend at a revolutionary war reenactment of the Battle of Germantown, in Philadelphia.

Here is the default rendering of the DualISO DNG, converted using the cr2hdr application - the DNG was rendered in RPP (Raw Photo Processor):

Input file     : _MG_0183.CR2
Canon EOS 5D Mark II detected
Full size      : 5792 x 3804
Active area    : 5634 x 3753
White level    : 12500
Black borders  : 158 left, 51 top
Black level    : 972
ISO pattern    : BddB GBRG
Noise levels   : 13.57 9.03 9.37 12.44 (14-bit)
Estimating ISO difference...
ISO difference : 4.21 EV (1845)
Black delta    : -9.86
Interpolation  : mean23-chroma5x5-alias
Dynamic range  : 10.32 (+) 9.73 => 13.94 EV (in theory)
Matching brightness...
Looking for hot/cold pixels...
Hot pixels     : 23
Full-res reconstruction...
ISO overlap    : 3.1 EV (approx)
Half-res blending...
Chroma filtering...
Building alias map...
Filtering alias map...
Smoothing alias map...
Noise level    : 2.96 (16-bit), ideally 2.51
Dynamic range  : 13.93 EV (cooked)
Black adjust   : -4
Output file    : _MG_0183.DNG

Here is a static example - I will go back and reshoot this scene when there is sun, and I will shoot it with and without the DualISO setup.

default rendering in RPP and a +4 exposure rendering in RPP. I find it better to demonstrate the 4 stop push with the precision of RPP compared to PS. Overall image and 100% crops ( I rendered at 1/2 res in RPP, so these are 50% crops, really).

cr2hdr log:

Input file     : _MG_0081.CR2
Canon EOS 5D Mark II detected
Full size      : 5792 x 3804
Active area    : 5634 x 3753
White level    : 15258
Black borders  : 158 left, 51 top
Black level    : 978
ISO pattern    : BddB GBRG
Noise levels   : 8.94 6.96 7.27 8.64 (14-bit)
Estimating ISO difference...
ISO difference : 3.17 EV (898)
Black delta    : 9.30
Interpolation  : mean23-chroma5x5-alias
Dynamic range  : 11.00 (+) 10.64 => 13.81 EV (in theory)
Matching brightness...
Looking for hot/cold pixels...
Hot pixels     : 396
Full-res reconstruction...
ISO overlap    : 4.8 EV (approx)
Half-res blending...
Chroma filtering...
Building alias map...
Filtering alias map...
Smoothing alias map...
Noise level    : 3.16 (16-bit), ideally 3.14
Dynamic range  : 14.14 EV (cooked)
Black adjust   : -3
Output file    : _MG_0081.DNG

Here is an example with a moving element. Same idea, original 0EV rendering from RPP, then +4EV. For the 100% crops i rendered at full res from RPP this time, so they are full res 100% crops.

cr2hdr log:

Input file     : _MG_0093.CR2
Canon EOS 5D Mark II detected
Full size      : 5792 x 3804
Active area    : 5634 x 3753
White level    : 15257
Black borders  : 158 left, 51 top
Black level    : 973
ISO pattern    : BddB GBRG
Noise levels   : 14.04 9.55 10.12 13.17 (14-bit)
Estimating ISO difference...
ISO difference : 4.21 EV (1850)
Black delta    : -9.47
Interpolation  : mean23-chroma5x5-alias
Dynamic range  : 10.55 (+) 9.99 => 14.20 EV (in theory)
Matching brightness...
Looking for hot/cold pixels...
Hot pixels     : 50
Full-res reconstruction...
ISO overlap    : 3.3 EV (approx)
Half-res blending...
Chroma filtering...
Building alias map...
Filtering alias map...
Smoothing alias map...
Noise level    : 4.89 (16-bit), ideally 2.61
Dynamic range  : 13.51 EV (cooked)
Black adjust   : -3
Output file    : _MG_0093.DNG

Pretty much like Zero Noise only done in your camera all at the same exact instant in time.

Totally bad ass.

kirk

A panning example! At full res in the deep shadows, when boosted, you can see the artifact of the line pairs. Not a biggie. At half res, they are gone.

cr2hdr log:

Input file     : _MG_0059.CR2
Canon EOS 5D Mark II detected
Full size      : 5792 x 3804
Active area    : 5634 x 3753
White level    : 15257
Black borders  : 158 left, 51 top
Black level    : 972
ISO pattern    : BddB GBRG
Noise levels   : 13.91 10.05 10.43 13.02 (14-bit)
Estimating ISO difference...
ISO difference : 4.22 EV (1859)
Black delta    : -7.10
Interpolation  : mean23-chroma5x5-alias
Dynamic range  : 10.47 (+) 10.00 => 14.22 EV (in theory)
Matching brightness...
Looking for hot/cold pixels...
Hot pixels     : 26
Full-res reconstruction...
ISO overlap    : 3.3 EV (approx)
Half-res blending...
Chroma filtering...
Building alias map...
Filtering alias map...
Smoothing alias map...
Noise level    : 7.37 (16-bit), ideally 2.58
Dynamic range  : 12.92 EV (cooked)
Black adjust   : -4
Output file    : _MG_0059.DNG

Another example - indoor-outdoor scene. RPP rendering at half res.

kirk

cr2hdr log:

Input file     : _UAL0034.CR2
Canon EOS 5D Mark II detected
Full size      : 5792 x 3804
Active area    : 5634 x 3753
White level    : 15259
Black borders  : 158 left, 51 top
Black level    : 972
ISO pattern    : BddB GBRG
Noise levels   : 10.71 5.94 6.01 10.49 (14-bit)
Estimating ISO difference...
ISO difference : 3.20 EV (918)
Black delta    : 25.09
Interpolation  : mean23-chroma5x5-alias
Dynamic range  : 11.23 (+) 10.38 => 13.58 EV (in theory)
Matching brightness...
Looking for hot/cold pixels...
Hot pixels     : 111
Cold pixels    : 17347
Full-res reconstruction...
ISO overlap    : 5.0 EV (approx)
Half-res blending...
Chroma filtering...
Building alias map...
Filtering alias map...
Smoothing alias map...
Noise level    : 3.76 (16-bit), ideally 3.76
Dynamic range  : 13.89 EV (cooked)
Black adjust   : -5
Output file    : _UAL0034.DNG

A side note that might be of interest to some for whom file size is a consideration. The dng file, apparently because it is 16 bit, is somewhat larger than the original CR2. 42 MB rather than the 22-30 MB of the CR2 (5D2). In LR (and I imagine in PSCS) the dng can be exported as a new dng - presumably with no loss of data - which is 25 MB. It may be that LR-produced dngs are more compressed than those from CR2hdr. There is also the option to export it as a lossy dng which would reduce file size further.

P.S. That CR2hdr does already do considerable compression may be judged from the fact that the size of an uncompressed 5D2 16 bit file would be 120 MB.

Edit: (Oct, 12) Dumb mistake. I forgot that the dng is greyscale and therefore would be 40 MB, so the CR2hdr output is uncompressed.

Here is a difficult scene to capture -perfect for a comparison between the ML dual ISO result and a standard raw capture!

Methods:

I used Raw Photo Processor (RPP) to perform the conversion of the DNG file generated from the ML data as well as the 5DII raw file of the scene. Both shots were acquired at the same exposure and WB settings and no NR was used other than the very limited RPP tools (remove hot pixels, blur chroma).

It is overcast and raining here, outside of Philadelphia - it has been for two days. This is a scene that may typically be approached with HDR imaging techniques. Outside there is a bright (not bright direct sunlight, but bright) scene, with an interior scene lit by a window. I shot this scene so that the manikin head was essentially backlit - the black desktop bounces very little fill into the face, but the white pad of paper to the right of the face does provide a somewhat concentrated bounce fill.

The goal of my rendering and edit:

I am sitting at my desk trying to match, visually, what I see in the real scene versus what my edit is portraying.

As I perform very basic edits, I am applying the edits to both images identically. This may not be the best strategy in real life, as you often have to use different edits to get the most out of an image, but in this exercise I think it is fair and will permit comparison of the two images.

I will not go into detail as to how I edited - I rendered a conservative TIFF from RPP:

Here is a composite screenshot of each of the two images in RPP, both boosted in exposure 4 stops.

Input file     : _MG_0001.CR2
Canon EOS 5D Mark II detected
Full size      : 5792 x 3804
Active area    : 5634 x 3753
White level    : 12500
Black borders  : 158 left, 51 top
Black level    : 974
ISO pattern    : BddB GBRG
Noise levels   : 12.75 9.05 9.22 12.13 (14-bit)
Estimating ISO difference...
ISO difference : 3.18 EV (904)
Black delta    : -9.59
Interpolation  : mean23-chroma5x5-alias
Dynamic range  : 10.31 (+) 9.82 => 13.00 EV (in theory)
Matching brightness...
Looking for hot/cold pixels...
Hot pixels     : 16
Cold pixels    : 1
Full-res reconstruction...
ISO overlap    : 4.1 EV (approx)
Half-res blending...
Chroma filtering...
Building alias map...
Filtering alias map...
Smoothing alias map...
Noise level    : 5.08 (16-bit), ideally 4.86
Dynamic range  : 13.15 EV (cooked)
Black adjust   : -7
Output file    : _MG_0001.DNG