Apertus Axiom Beta

Started by Andy600, May 09, 2014, 01:02:28 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.


Kharak

Aren't they way behind on schedule?

I thought the first betas were supposed to be released Q1/2 of this year?
once you go raw you never go back

LightCtrl

Quote from: Kharak on December 04, 2015, 11:21:48 AM
Aren't they way behind on schedule?
I don't know exactly how much, but yes they are.   To be fair some delays were expected and they added extra functionality/flexibility in the hardware.

First sequences from the prototype with finalized hardware are up (minus full bandwidth sensor board with fpga).  You can see them in the latest update video.
https://www.youtube.com/watch?v=izDzFov3nDM&feature=youtu.be

You can also download one clip as a 4096x1080 12bit DNG sequence: http://files.apertus.org/AXIOM-Beta/coffee_4K_raw_highspeed_burst.zip

My layman's impression (ignoring some of the artifacts that will be addressed) is the image detail is great, though that might be expected from 12bit raw in a studio, However the FPN ruins the experience (its not yet corrected).  The footage can be pushed quite a bit.  I found a few looks I liked with it.

My question for @a1ex , Sebastian or anyone else who knows their shit is:  How much of the FPN can usually be corrected?  Is the betas implementation likely to offer the same noise performance as BMPC or AJA cion, where FPN rears its ugly head at iso800 and above, or can this be improved upon? Also does the original assessment of DR hold true with the new footage.

Back when I purchased my voucher I was hoping this would be an alternative to other high fps cams I could never afford, and as a huge fan of Ptools and ML it seemed like a natural fit.  Since, my work has evolved in the year+ since the campaign and I now have a FS700, another $2k for just a fun project seems excessive.  Given the footage from the BMPC and Cion its hard to see the Beta w/ cmv1200 touch the FS700 in sensitivity and DR (or even really match it).  Plus in its current form you can't record from it without a pc. 

I would like to see internal rec to SD, or an output that works with available external recorders.  I would like to see the full sensor readout with a usable duration: 5-10sec.

I really wan't this cam to work out, and put more control in my hands then I've ever had before.  I would like to continue to support the Axiom team, I think its great what they are doing and have eagerly awaited each update.   However I can't justify spending that kind of money unless I can get great images from it and have started to worry that the camera might not be for me.





a1ex

Quote from: LightCtrl on December 04, 2015, 05:30:41 PM
My question for @a1ex , Sebastian or anyone else who knows their shit is:  How much of the FPN can usually be corrected?

A while ago I've tweaked a few DNGs, and you can find the result at the bottom of this page: https://apertus.org/axiom-beta-hello-world-article-may-2015

Here's a more recent sample (DNG available on request), where I tried to find a color matrix. Left is uncorrected, right is corrected. There are some tearing artifacts that you can safely ignore.



How much of this will be fixable in real-time and how much will be postprocessing... we'll see. I haven't really started working on it, but it will probably happen soon.

Quote
Is the betas implementation likely to offer the same noise performance as BMPC or AJA cion, where FPN rears its ugly head at iso800 and above, or can this be improved upon?

I haven't seen samples from these cameras yet, so if you have some DNG files, I can take a look (and maybe even use them as test data for my algorithm, if the noise profile is similar).

QuoteAlso does the original assessment of DR hold true with the new footage.

The latest test clips were taken with crushed blacks (wrong sensor configuration), so unfortunately I couldn't recover anything from the shadows. Hopefully the next samples will match my measurements.

LightCtrl

A1ex, thanks for your answers. 

I must have missed your tweaked DNG's the first time around.  Much improved, and way better then what I got myself.   

Here is a link to BMPC footage.  Hope it helps.  While significantly better then anything we've seen from the beta its still seems pretty noisy.  I suppose that sort low backlit scene would be pretty noisy with a fair number of cameras.

https://mega.nz/#!FJpBzRZK!7gcUhMcmivnpsgIisyO_-oPufMYSVYzw13OBykd1Cec

Bm rates the 4k production camera at 12 stops usable is probably around 8, which is gh2 territory.  Hopefully the beta can do better.  It will have different processing, and maybe some of the magic sauce A1ex mentioned in the beginning of this thread works and it can at least get up to canon level.

a1ex

Do you mind extracting a few DNG frames for me? I'm on a very slow network connection.


a1ex

I don't see any pattern noise on the first DNG (and I assume the others are from the same movie). Am I missing something?

LightCtrl

No I don't think you are.  I didn't realize you were looking for shots with FPN specifically.  I just see some nasty looking color noise in the dark shadows (waves and rocks).

Here are some DNG's that defiantly have FPN.  Their from the Ursa mini but the 4k version that by all accounts has the cmv12000 chip.

https://drive.google.com/folderview?id=0ByDzUJpFXhFjS0RTZ2F0eUhtcFE&usp=sharing

LightCtrl

According to Twitter the Axiom team has made some progress with image processing already.
https://twitter.com/ApertusOSCinema/status/674172374376390656/photo/1?ref_src=twsrc%5Etfw

No samples available yet.  I am highly interested in the results.

a1ex

These were exactly what I was looking for, thanks. I have finally managed to fix those without introducing severe artifacts, and here are the corrected files:

https://www.dropbox.com/sh/izyk4mn29429do9/AABDPunduQWSntQKJTiDFlbxa?dl=0

Here's a half-res jpeg, before and after:


And a 100% crop from a full-res jpeg (rendered with RawTherapee):


You will still have to use green equilibration or the VNG4 algorithm in your raw processing software (I didn't correct this artifact yet).

The algorithm still gets confused in some cases (e.g. with strong horizontal or vertical edges in the image), but I'd say it's pretty usable. It will probably be helpful on Canons as well, since some models suffer from FPN at certain ISOs.

Source code: patternnoise.c

Anyone interested in a FPN correction tool for any other camera? (be it Canon, BlackMagic or whatever else, as long as dcraw opens its raw files)

hjfilmspeed

Yes!!!! 5d3 CR2s or MLV/RAW pretty please! This looks great to me!

Danne


jimiz

CANON 5d3 please RAW & MLV , thanks!!!
5D3-123

Kharak

What do you guys mean about the 5D3?

A1ex made the Vertical Stipe fix a loooong time ago.

@A1ex, amazing work.. You are the Savior of Cameras.
once you go raw you never go back

dmilligan

FPN and "Vertical Stripes" are two very different issues.

If I have time, I may implement it in MLVFS. Shouldn't be too difficult.

Kharak

I was under the assumption that the Vertical Stripes are there because of the Fixed Pattern Noise from the sensor.


What are the vertical stripes then?
once you go raw you never go back


a1ex

Or better, http://theory.uchicago.edu/~ejm/pix/20d/tests/noise/#patternnoise

The pattern noise we are seeing in the above images is not exactly fixed (it varies from frame to frame), and it's a scalar offset applied to each row, and then the same process is repeated for each column. So, to correct a 4096x3072 image, you will need 7168 scalars. In the Apertus sensor, I believe things are a little more complex than that (the offset seems to be dependent on image brightness, but I didn't find a clear relationship yet, and the constant offset model already gives pretty good results).

In 5D3, vertical stripes appear because each column ( modulo 8 ) has a different gain, so the artifact is visible only in highlights. That is, columns 1, 9, 17 and so on have one gain, columns 2, 10, 18 ... have some other gain, so in order to correct this artifact, one has to compute 8 gains (scalars). I believe this happens because the sensor is read out using 8 different circuits in parallel, each with slightly different electrical parameters.

Other cameras: I believe 6D and 700D use 4 circuits, 60D/600D/550D/5D2/50D use 2 circuits, and 500D uses only one circuit. I came to this conclusion by checking the values of the sensor resolution registers (vertical resolution is always unscaled, while horizontal resolution is divided by 8, 4, 2 or 1, depending on the camera). Of course, this parallel readout affects the pattern noise as well (for example, on 60D, if you take the uncorrected*) pattern noise from even columns, and then the one from odd columns, you get two independent random walks).

*) To completely disable Canon's pattern noise correction, one has to replace all the W+H offsets with 0 (they are passed via EDMAC to the HivShd module).

Hot pixels are another story, and they are usually easy to find and fix.

Kharak

Thanks a lot for explaining A1ex.

Understanding every element of my preferred cameras' makes me a better Video/cinematographer. In knowing why and what is causing these things, I feel that the way I capture my images is on a much more intricate level, from Planing to motive to subject to framing to lens to sensor to internals to data to transfer to Processing to Post Processing to Cutting to a finished product. Basically, I like knowing why!

once you go raw you never go back

dmilligan

Quote from: dmilligan on December 29, 2015, 05:58:55 PM
If I have time, I may implement it in MLVFS. Shouldn't be too difficult.
Done

kgv5

Was this update also implemented in PC version of MLVFS?
www.pilotmovies.pl   5D Mark III, 6D, 550D

Licaon_Kter


Quote from: kgv5 on January 04, 2016, 03:32:24 PM
Was this update also implemented in PC version of MLVFS?
If you compile the DLL I guess it is.


Does that work ok anymore?

dmilligan

Quote from: Licaon_Kter on January 04, 2016, 03:45:23 PM
If you compile the DLL I guess it is.


Does that work ok anymore?
No. The PFMAP code has not been updated in a very long time and is certainly broken by now.

jtvision

Quote from: a1ex on December 14, 2015, 08:55:39 PM
Anyone interested in a FPN correction tool for any other camera? (be it Canon, BlackMagic or whatever else, as long as dcraw opens its raw files)

Yes, please! Here's the link to the noisy .dng files from Blackmagic Production Camera 4K. Thank you!

https://drive.google.com/folderview?id=0B2KgeL11wjTZZlNTTDRYQjJTdnM&usp=sharing