Dual ISO - massive dynamic range improvement (dual_iso.mo)

Started by a1ex, July 16, 2013, 06:33:50 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

a1ex

Just found a bug in the interpolation code near clipped highlights. Updated converters in first post.

Before (half-res ISO 1600 interpolated):


After (half-res ISO 1600 interpolated):


Please remember that in video mode it's skipping 6 lines at a time, and in photo mode it's skipping only 2 at a time. So my suggestion is to try to post photo samples too; videos are not the strong point for this method.


stevefal

The artifact you fixed looks a lot like the "serrated" edges in the video in the previous post. In light of your note regarding video versus photo use of the feature, will this fix improve the video case nonetheless?
Steve Falcon

Audionut

Quote from: horshack on July 17, 2013, 10:33:11 PM
The ISO 100 lines will have more read noise and tossing them out will improve DR but at the same time will reduce photons collected by half, reducing the shot-noise SNR of those shadows from 22.93:1 to 16.21:1

Can you explain the relationship between resolution and SNR?

dubzeebass

I just did a test using fine details (fur and grass) to demonstrate the line-skipping.  This is of course a "worst-case" video, but it should help to show the interlacing.

Dual-ISO: http://youtu.be/w4PxA1KcUFA
Single-ISO: http://youtu.be/yiMdOVfCgJ0

Marsu42

As with a lot of recent new features I'm wondering "Why should I care, will this eventually available on the xyz (replace with your camera model)?"

It'd be great if this information would make it in news announcements w/o the need to flip through dozens of pages, because you never know if some features *require* a dual-digic or new camera (like dual iso 7d/5d3) or if these are just the first models that get it.

ShootingStars

I'm sick and tired of these "wish it was on XYZ model" posts.

Audionut

Quote from: Marsu42 on July 18, 2013, 02:56:50 AM
It'd be great if this information would make it in news announcements w/o the need to flip through dozens of pages, because you never know if some features *require* a dual-digic or new camera (like dual iso 7d/5d3) or if these are just the first models that get it.

The information is linked in the pdf in the first post.

Asking people to read is not an overly unreasonable expectation imo.

1%


/* 00:00:02.660613  */ CMOS
/* 00:00:02.660658  */     00 0593 404520BC
/* 00:00:02.660702  */     01 0000 404520BE
/* 00:00:02.660742  */     02 0225 404520C0
/* 00:00:02.660782  */     03 0000 404520C2
/* 00:00:02.660820  */     04 04F0 404520C4
/* 00:00:02.660859  */     05 0433 404520C6
/* 00:00:02.660898  */     06 0000 404520C8
/* 00:00:02.660937  */     07 0000 404520CA
/* 00:00:02.660974  */     08 000A 404520CC


CMOS regs for 6D I hope... maybe this will work for ADTG too.

It did! ADTG shutters for all!

Marsu42

Quote from: Audionut on July 18, 2013, 03:12:29 AM
The information is linked in the pdf in the first post. Asking people to read is not an overly unreasonable expectation imo.

Thanks for the information, given with the usual finesse if I may be so bold to comment. I'd even assist you in your reading efforts: the pdf is linked with the description "in-depth description of how it works" which is exactly not where I'd expect to find the very basic information "will this come to my camera" - that's why I've suggested to provide this in a more obvious way, i.e. neither buried in a pdf nor somewhere in the thread.

For everyone else: No, it's not coming to cameras except 7d/5d3 :-\

1%

QuoteNo, it's not coming to cameras except 7d/5d3

Lol, don't be so sure.

driftwood

Dual ISO artefacts with mean23 but much less than with standard current raw2dng.

Canon 60D, Canon 5DMK3, Lexar 1000x 128GB CF, Panasonic (shhhh!) GH2s & GH3s. :-)

squig

Quote from: driftwood on July 18, 2013, 04:56:12 AM
Dual ISO artefacts with mean23 but much less than with standard current raw2dng.

Wow!

Did you get a 100/1600 shot?

Ryan Lightbourn

Quote from: NateVolk on July 17, 2013, 09:08:53 PM
This is great alex!  I'm happy to help test with my 5d3, but I'm a mac guy and having some difficulties post processing the files.  Read the pdf, downloaded the zips from page one, but then what?  Sorry for the stupid ?'s...
I searched for a tutorial on using the cr2hdr.exe and the draw.exe but no luck.  Thanks anyone for some direction!

2nd that...also on Mac.  Got some shots, just can't figure out how to make them look normal in AE.
twitter.com/ryanlightbourn
instagram.com/ryanlightbourn
ryanlightbourn.net

a1ex

@driftwood,

Can you upload 1 second (or less) of raw video (the RAW file, not the DNGs), where mean23 gives better results than the default median?

In all my tests, median was a lot better on edges (see my bugfix above - that's an interpolated image, 2 lines from the sensor, 2 interpolated). I did not shoot power lines, I don't like them :P

squig

Quote from: a1ex on July 18, 2013, 08:54:01 AM
I did not shoot power lines, I don't like them :P

Who does?

From what I've seen so far 100/400 has a lot more DR, has Samuel run his DR test?

Iain

I hope this is helpful, sorry if it is not.

I downloaded the CR2 and DNG to have a play with and I found that by denoising the noisy rows in the shadow areas before demosiacing and then demosicing all rows some artefacts are reduced.

This is my version created from the cr2 (via dcraw then G'MIC) and then the DNG version (via UFRAW).




The denoising is a simple median filter on the noisy rows. It is applied to 3 consecutive pixels of the same colour EG R1,R2 and R3 and also on G1,G2 and G3 and so on.

R1 G1 R2 G2 R3 G3

I repeated the filter 3 times.

I used my own demosaicing method. I'm not a programmer, I just play with scripting in G'MIC , but I thought this might be useful.

araucaria

Quote from: 1% on July 18, 2013, 03:54:08 AM
Lol, don't be so sure.
That would be great!

Anyway, what happened to iso-less? The increased DR is great for the guys tied to canon but the really interesting thing was that iso-less thing.

hetfanatic

I'd like to test these for 7d but I'm still not sure how to implement / install it :(

Stedda

You need to compile the code for yourself at this point... there are tutorials here on the site.

or you must wait for either a forum user to host their compiled code of the Devs to release an updated Alpha.
5D Mark III -- 7D   SOLD -- EOS M 22mm 18-55mm STM -- Fuji X-T1 18-55 F2.8-F4 & 35 F1.4
Canon Glass   100L F2.8 IS -- 70-200L F4 -- 135L F2 -- 85 F1.8 -- 17-40L --  40 F2.8 -- 35 F2 IS  Sigma Glass  120-300 F2.8 OS -- 50 F1.4 -- 85 F1.4  Tamron Glass   24-70 2.8 VC   600EX-RT X3

a1ex


Seanc

I prefer 'b' because it has more texture detail than 'c', and tames the chroma noise that's objectionable in 'a'.

horshack

Quote from: sarangiman on July 17, 2013, 10:45:50 PM
Hi horshack-- why would you 'reduce photons collected by half'? The exposure at the sensor plane is not changed; there is only one exposure. The ML hack is simply changing what you do with that exposure data, & so should not affect the SNR save for actually increasing it in shadows by reducing the effects of downstream noise (after ISO amplifier) in the ISO 1600 exposure.

The focal plane exposure (in terms of shutter speed/aperture) remains the same for both ISO 100 & ISO 1600 exposures; it's not like two separate exposure are being made (where the ISO 1600 exposure would have 4 stops less exposure which, yes, would increase the effects of shot noise).

Let me know if I'm misunderstanding you.

For the shadows the ISO 100 lines are being discarded from the composite raw, which reduces the light-capturing surface area of the sensor in half.

stevefal

I'd choose c. The first two have too much chroma noise in the shadows. I think texture detail between b and c are comparable.
Steve Falcon

driftwood

a1ex: Not at home right now so can't send you original raws. Here's my latest screenshots shoot produced today showing the processing of the same RAW file fed thru each of the methods: mean23 v median6 v shadow100.

https://www.dropbox.com/sh/4rpnyayri6n1jrk/2iMOz1n5Nl/mean2rb3g%20450%20avg%20hot%20pixels%20per%20frame.jpg

https://www.dropbox.com/sh/4rpnyayri6n1jrk/osCxpEq1xz/median6-1500%20avg%20frame%20hot%20pixels.jpg

mean23 averaging 450 hot pixels per frame.

median6 1500 per frame.

Shadow100 averages 1000 hot pixels per frame on the same tested RAW file.

EDIT: UPDATED (now with shadow100) picture showing the major differences of each against each other:
https://www.dropbox.com/s/2h7x59girnbxdpu/mean23_median6_shadow100.png

As to your A,B,C, question the grain in A to C goes from courser to finer chroma noise. Most people are going to choose the aesthetically more pleasing C, but I'd probably hunch for B. There's a tad more shadow detail in A which could be processed in post.
Canon 60D, Canon 5DMK3, Lexar 1000x 128GB CF, Panasonic (shhhh!) GH2s & GH3s. :-)

a1ex

Here's how I've got the 3 pictures:

a) should have no shadow aliasing (shadows were handled as if they were midones). To get similar results, use this raw2dng: http://acoutts.com/a1ex/raw2dng_shadow100.exe

b) is a 50% blend between a) and the high ISO image. Less noise, a little aliasing. http://acoutts.com/a1ex/raw2dng_shadow050.exe

c) is 25% a) and 75% high-iso. Therefore, even less noise, but a little more aliasing. http://acoutts.com/a1ex/raw2dng_shadow025.exe

By default, shadows used data from high ISO only. This gets minimal noise, but it seems to have serious problems with aliasing.

Of course, now the formula from Horshack's friend should be valid for case a).

@MA Visuals: can you repeat your test with these 3 raw2dng's? Just post-process the same video with them.

Latest cr2hdr uses 50% blend (link in first post).

@ajay and @audionut: can you develop your test shots (the same pictures that you posted earlier) with latest cr2hdr?