The image is underexposed by more than 3 stops.
Based on
http://www.magiclantern.fm/forum/index.php?topic=10111.msg118553#msg118553, my advice would be to use a slightly higher exposure time (maybe 1/33 - that one shouldn't cause trouble with artificial lights, just like 1/50) and increase ISO to 3200 or even 6400. On 5D3, the benefits of increasing ISO are much higher in regular 1080p or 720p video modes, compared to photo mode or 1:1 crop, because of... pixel binning.
1080p/720p: by increasing ISO from 3200 to 6400, you lose 1 stop of highlights to gain about 0.5 stops in shadows. Above ISO 6400, there's nothing more to gain.
1:1 crop / photo mode: by increasing ISO from 1600 to 6400, you lose 2 stops of highlights to gain about 0.5 stops (0.36+0.24) in shadows. You'll probably think twice before increasing ISO above 1600 or 3200.
# compute shadow improvement when increasing ISO (100->200, 200->400, ..., 6400->12800)
# in other words: how cleaner your shadows will be after increasing ISO by 1 stop (thus clipping 1 stop of highlights)
# measurements done on 5D3, http://www.magiclantern.fm/forum/index.php?topic=10111.msg118553#msg118553
# isos = [ 100 200 400 800 1600 3200 6400 12800 ];
octave:1> dr_1080 = [ 11.10 11.05 11.02 10.93 10.73 10.39 9.88 8.88 ];
octave:2> dr_crop = [ 11.22 11.13 11.00 10.82 10.32 9.68 8.92 7.70 ];
octave:3> 1 + diff(dr_1080)
0.950 0.970 0.910 0.800 0.660 0.490 0.00
octave:4> 1 + diff(dr_crop)
0.910 0.870 0.820 0.500 0.360 0.240 -0.22
# note: output aligned manually for easier reading
In your case: 1/50 -> 1/33 means more photons captured (0.6 EV more), and ISO 800 -> 6400 will improve the shadows by about 2 stops (0.8+0.66+0.5), while clipping 3 stops of highlights. In your test image, clipping 3 + 0.6 EV of highlights would affect a grand total of... 72 pixels (0.0035% of the image area).
Math:
dcraw -4 -E 1Z7A2572_000422.dng
octave:1> a = imread('1Z7A2572_000422.pgm');
octave:2> prctile(a(:), [50 90 99 99.9 99.99 99.999 100])'
ans =
2094.0 2184.0 2306.0 2402.0 3037.0 3335.5 3493.0
# overexposure amount (using default white level from MLV metadata)
octave:3> log2(16200-2048) - log2(3493-2048)
ans = 3.2919
# how many pixels would be clipped after increasing the exposure by 3.6 stops?
octave:4> 2 ^ (log2(16200-2048) - 3.6) + 2048
ans = 3215.1
octave:5> sum(a(:) > 3215)
ans = 72
The algorithm used for fixed pattern noise has a major weakness: it gets easily confused by horizontal or vertical lines in the actual image. I have some ideas to rework the algorithm, by looking at past and future frames and applying some sort of optical flow, but it's not easy.
A dark frame may improve things, but
my previous attempt in a similar case was unsuccessful (because the pattern is not exactly repeatable with a dark frame - instead, my understanding is that Canon code fine-tunes it continuously, with a time constant of a few seconds - see e.g. the vertical artifacts in centered x10 zoom with crop_rec_4k builds).
I'd like to take a closer look at this pattern change, but cannot promise a timely fix. Do you mind uploading a sample MLV?
