Night BW shot using old LX7

heckflosse · January 8, 2021, 9:27pm

Enough of ot now. Let’s continue play raw

heckflosse · January 8, 2021, 9:34pm

heckflosse · January 8, 2021, 9:52pm

kofa · January 8, 2021, 10:14pm

Very nice, though I prefer the softer reflections from the camera’s version (and also @Phil_Smith’s, @elstoc’s, @age’s BW).
@Jade_NL: I was an RT user before darktable. Actually, I started using darktable because of its OpenCL acceleration (my Core2 Duo with 4 GB is not quite a powerhouse; will perhaps replace it this spring). But I do keep an eye on RT & ART, and Filmulator, too.

priort · January 9, 2021, 12:07am

Is RCD in the master branch?? I did a build a couple of days ago and I don’t have it??

pphoto · January 9, 2021, 12:47am

Nice image. I decided to stay in a colored version although I had to struggle a bit with the various colored lights.

P1050600.RW2.xmp (11.5 KB)

kofa · January 9, 2021, 8:49am

github.com/darktable-org/darktable

Introducing a tiled RCD demosaicer (#7337)

committed 09:36AM - 29 Dec 20 UTC

jenshannoschwalm

+698 -51

* Introducing a tiled RCD demosaicer The rcd demosaicer developed by Luis San…z Rodríguez is publically available at https://github.com/LuisSR/RCD-Demosaicing and is licenced under the GNU GPL version 3. I came across this via the rt issue tracker Beep6581/RawTherapee#5908 and have personally been using it in dt for some months. The results are very good, quality is at least en-par with Amaze, maybe depending on image content the results are slightly better or slightly worse. The biggest pro argument is less color overshoot, the penalty is marginally increased blur. Having a demosaicer output with less overshoot makes this a better option for further processing. I started developing the demosaicer by taking the algo from original github and adopted it to dt style. Over time i learned how @heckflosse had improved the code by tiling, less omp overhead and by adding sse2 code in the most time-critical internal loop as the compiler does not vectorize here as good. Further on we discussed many details and worked on performance on different machines. Thanks a lot! Now it is really good, on darktable rcd is ~3x faster than amaze and 30% faster than the non-cl ppg code. On my machine (pretty similar to the well known 9900k) i am now at ~135ms for a 45Mpix image. This is ~3 times as good as the non-tiled version. The border interpolation used the dcraw algorithm in the original code, there are minor issues so i chose the code used by rt as it has less artefacts. A border of 9 is always safe if tiling is used. Unfortunately there is no paper describing the algorithm, we only have the README in the original git repo at https://github.com/LuisSR/RCD-Demosaicing/blob/master/README.md Also there is no formal quality analysis, at rawtherapee we have https://rawpedia.rawtherapee.com/Demosaicing But we have a lot of experience people made when using this demosaicer in rt. Some notes for people testing the code 1) in demosaic.c you find the bitmask #define DEMOSAIC_TIMING 0 - 01 --> add timing info for the demosaicers while tuning avoiding the mass of -d perf info - 02 --> prohibit fall back to ppg 2) You might want to tune the tiling size in rcd_demosaic.c, unfortunately this has to be selected at compile-time. Making this runtime specified reduces performance by ~50% See former attempt #6738 * Add proper #include for optional sse4.1 intrinsics * Work on the SSE specific code sections 1. #ifdef the sse specific macros/inlines to avoid rcd being sse specif code. 2. The indx>>1 trick had been used to lower cache-mem pressure but lead to bad compiler vectoring. All indexes to lpf now use an extra index to avoid the shifting. (2) improves the non-sse performance by ~20% on my machine but still there is a significant gain of overall ~4% if using the hand-optimized code so i would prefer to keep this. * Fixing corner cases with border data * Using index/2 also for PQ_Dir As in steps 2.1 and 3.1 we can use index/2 as we calculate data only for every second pixel in a row. Reduces cache mem pressure and thus increases performance. * allocate tiling memory by dt_alloc_align_float As @ralfbrown suggests in #7419 comments and elsewhere we should always use dt_alloc_align_float instead of dt_alloc_align to get rid of the '64' magic. * Code maintaining keeping track of common/math.h The CLIP macro is only in 3.6 branch so we define it in rcd demosaicing if undefined. * Fixing compiler issue after common/math.h work 1. The CLIP macro is not necessary any more as defined in common.math.h 2. float (*const rgb)[RCD_TILESIZE * RCD_TILESIZE] = (void *)dt_alloc_align_float needs a cast

hannoschwalm · January 9, 2021, 4:42pm

It is in master branch but unfortunately i introduced a bug while optimizing performance. There is a pr fixing this so maybe you wait for a day until it gets merged

kofa · January 9, 2021, 5:01pm

Thanks for the quick fix. I hope to test this tonight, once I get home.

Jean-Marc_Digne · January 9, 2021, 5:25pm

thanks for posting, I have to practice night shots with DT 3.4

P1050600_02.RW2.xmp (22.6 KB)

Suki2019 · January 31, 2021, 9:39am

this is a real challenge to reproduce this SW look

that’s closest to what I can do with tonecurve on my laptop.
maybe I’ll try again on my real monitor.

P1050600.RW2.xmp (10.1 KB)

kofa · January 31, 2021, 2:18pm

Are you sure that’s the right sidecar? The original is a Panasonic RW2 file

Suki2019 · January 31, 2021, 2:49pm

damned…

is changed

kofa · January 31, 2021, 5:14pm

Thanks, everyone. By chance, I opened this file again today, before I noticed the new entries.
With the colour version, my experience is that the plate’s background tends to turn into pale blue (I got results similar to @pphoto Night BW shot using old LX7 - #54 by pphoto), when in reality it’s a kind of green (as shown in Night BW shot using old LX7 - #26 by priort, Night BW shot using old LX7 - #29 by priort and
Night BW shot using old LX7 - #31 by Underexposed).
The Yamaha sign is a pain, too, my own attempts did not get even close to what you’ve achieved.
Also, when I reduce the brightness of the plate, I can’t get a good mask, mine are all soft:

I’ll have to study your sidecars more closely.

kofa · February 1, 2021, 6:17am

BTW, this is literally 2 clicks in Filmulator:

B&W on
highlight recovery = 2

Too bad that there are artefacts after turning on highlight recovery. Without that, it’s the traffic light that looks bad.

With HL recovery:

Without:

@CarVac, any suggestions?

Suki2019 · February 1, 2021, 12:51pm

one more try on my monitor.

so I don’t get any closer.
the lack of details may also be a kind of tone mapping in the camera, which darktable cannot reproduce.
Nikon pictures ooc also have more details / microcontrast, and apart from nikon software itself, as far as I know, no other converte can reproduce this. the noise reduction engine in the camera also seems to be better optimized.

translated by Google

Night BW shot using old LX7_P1050600_01.RW2.xmp (9.9 KB)

afre · February 1, 2021, 1:21pm

This image is very interesting. Simple with a lot going on. None of the results match my visualization () of the actual scene yet. As they say, the night is young.

Suki2019 · February 1, 2021, 2:03pm

yes, … it is difficult to emulate / simulate or recreate something if you don’t know the basics of the process.

I’m not really satisfied with my attempts either.

translated by Google

CarVac · February 1, 2021, 2:24pm

If you take a look at the color image before you enable highlight recovery, you’ll notice that the issue is the sensor doesn’t clip very nicely, leaving artifacts at the edges of the lights. Could be partly due to sensor readout quirks (like the ol’ 5D2 black dots) or due to lens purple fringing.

Regardless of what causes the artifacts, this really messes up the highlight reconstruction.

One way you can deal with this in the monochrome conversion is to set highlight recovery to 1 (no clipping, no reconstruction) and then raise the red weight to keep the pink highlights from appearing gray. But this changes the look of things.

Suki2019 · February 1, 2021, 2:48pm

i have an idea …
maybe someone has the possibility to photograph a colorchecker with this b / w mode?
and can he provide the jpg and the raw file?
this should at least make it easy to reproduce the tone curve in DT.

I think it’s worth a try.

translated by …