Danne's crop_rec_4k experiments for EOS M

Started by Danne, December 03, 2018, 06:10:17 PM

Previous topic - Next topic

0 Members and 3 Guests are viewing this topic.

julien becker

Thanks Walter ! Now I get it. Indeed in play mode I got 94MB/s  :)

photomotion

Quote from: cedricp on September 20, 2022, 10:24:41 AM
Awesome, tested with wide lens (and speed booster), it's amazing. Using Resolve AI upscaler add even more sharpness to the image.
Great job dudes  8)
I'm wondering what we can expect more from that little beast

May I ask how you scale this in Resolve? In clip attributes, I can only find an option to scale the entire clip. How did you scale height/width separately?

Edit:
I want to export my clip as CinemaDNG. So, I can't use the options in transformation section of the MLV app. In resolve, the closest option I can think of is using a 2.39:1 timeline and selecting the stretch option under Retime and Scaling section.

gabriielangel

Follow up on the Frame Corruption discussion started here: https://www.magiclantern.fm/forum/index.php?topic=25841.msg240550#msg240550

The problem appears to be that the extra details introduced by focusing on the LCD screen at ISO 400 require more bandwidth than the card can handle.

Details:

I first tried modifying reg_713c and reg_7150 by +-10 each and compared the recorded file sizes to those recorded with the standard preset. Differences are marginal.
So those regs are probably timing related (High values distort the preview). I will test more in the coming days.

I then recorded files slightly soft-focused, and then focused a hair soft right before it breaks. There is a  27.9 MB/s difference between soft and almost full-focus.

So I recorded more files, starting a tiny bit soft focused and getting closer to full focus until it gave the corrupted frames error.
The closest file I got to full focus averages 94.7 MB/s, which is already close or above the benchmarked speed of the card (Sandisk Extreme Pro 170 128GB, 92.8-95.7MB/s)

So the problem appears to be that the extra details introduced by focusing on the LCD screen at ISO 400 require more bandwidth than the card can handle.

For the tests, I did a MLV fast-pass export of frames 25 to 96 (72 frames or 3 seconds) for each file I recorded, to facilitate comparison (Divide the file size by 3 to get MB/s):

Corruption-Test-Folders" border="0

MLV-Files" border="0

This screenshot is for the folder where I MLV fast-passed 3 seconds for each file, for the focus sharper until it breaks test:

Progressive-Focus" border="0



Danne


gabriielangel

If I set the camera right on the focus point where it breaks on demand, and then switch presets:

2.5k 14bits and 12bits : Breaks
2.5k 10bits : Record times are similar to a very exposed image, it just stops recording after a while, without any corruption.

2.8k 14bits  : Breaks
2.8k 10bits and 12bits : Record times are similar to a very exposed image, it just stops recording after a while, without any corruption.

5k frtp 10bits and 12 bits : Seems to be unaffected by this issue.
5k frtp 14 bits : Breaks

1080p 10bits and 12bits : Seems to be unaffected by this issue.
1080p 14 bits : Breaks

If we look at my previous tests in an informal manner, and add 27.9MB/s (The additional rate for a "noisy" picture) to the data rate figures:

https://www.magiclantern.fm/forum/proxy.php?request=https%3A%2F%2Fi.ibb.co%2FZGHKYLw%2FSandisk-Extreme-SDSQXA1-256-G-GN6-MA-Report2.png&hash=e09c4547a1295a140c4a23cf1a2f91e9

(I didn't publish it, but 5k frtp is 71.7 and 60.2 MB/s for 14 and 12 bits respectively.)

We get one more hint that maybe too much data is being sent to the codec or that the way the codec encodes the extra information yields more data than the card can handle?

MWilcut

I've done *something* and I can't figure out what or how to reverse.

ML was working just fine, no issues.  All of a sudden, when I start the EOS M and longpress the Trash icon, it does not load ML anymore.  I have uninstalled ML and reinstalled ML (even updating the ML to a more recent version of Danne's build).  I have reset all camera functions in photo mode...

What else can I try?

Walter Schulz

Quote from: MWilcut on September 24, 2022, 09:35:47 PM
I've done *something* and I can't figure out what or how to reverse.

ML was working just fine, no issues.  All of a sudden, when I start the EOS M and longpress the Trash icon, it does not load ML anymore.  I have uninstalled ML and reinstalled ML (even updating the ML to a more recent version of Danne's build).  I have reset all camera functions in photo mode...

What else can I try?

Uninstalling ML doesn't reset ML settings.
Delete ML\Settings contents and copy build files to card once more.

gabriielangel

Data Rate Upper Limits Warning: Long Post Ahead :)

Like I said in the previous post, I would like to create a test which can be easily and quickly replicated by anyone, in order to have a coherent comparison between the performances of various cards at different resolutions.

Because filming the LCD Screen leads to moiré artifacts which biases the results, I went outside to film various scenes at 2.5K to get a baseline. I will then create a pattern which can be printed on plain paper and filmed with the camera, with enough details to generate files of comparable size to real life recordings.

This test will also check where the "Data Rate Overload" occurs in real life (See my previous post)

The motivation for going through this is as follows:

1- I just purchased a high-capacity 200MB/s card to test the upper limits, which cost me half the price of the Camera.
2- The Students I help would rather spend that money on beer...
3- It took me 4 months of reading and testing before being able to use that EOS M to its full potential. This negates any low-cost benefit afforded by the cam. But if there was a set of guidelines written in a few posts, newcomers would be able to use the cam to its full potential in a fraction of the time (And with cheaper cards, if necessary).
4-It's always nice to have a list of several alternative cards. For example, the latest Sandisk Extreme Pro did not work well until the latest Overclock module.

Short Answer:
1- A test pattern which can generate 70-80MB/s files at the target exposure and 90MB/s+ at ettr would simulate real-life conditions.
2- The Bench.mo would be a lot closer to real-life use if it also ran a sustained-write test (1-2 min?)

3- With 240MHz Overclock, there is no need to heavily underexpose the image to get decent recording times at 12bits, and 14bits with little restrictions (2.5K tested here, but theoretically applies to 5k frtp and 1080p).

I avoid recording at 10 bits, as it gives a reddish tint to the noise in the image, and more noise overall. So 10bits should be carefully exposed to minimize the noise, and 12/14 bits should be preferred if there are lots of dark regions in the image.

Look at scene 5 below to see a real-life example of exposure triggering a frame corruption message.

Details:
Exposure being a key factor in achieving near-continuous recordings (Or long enough clips), let's establish what I consider a bright-enough, well exposed image. In the examples, I also included graded versions, so that it shows that the tests are not unreasonably brightly exposed.

I had the scopes in my MLV App modified to make it easier to evaluate the image. I left those in every screenshot, so that it shows that no clipping is taking place. (Click or download images to enlarge)

00-Mod-Scope" border="0

ScopesLabel.cpp
00-Scopes-Label-cpp" border="0

A few examples where I exposed with histogram at 1.0 and the main subjects Grey and Green with hints of yellow, when using the False Colors:

Example 1 Before and After Grading:
00-Well-Exposed1-Before" border="0

00-Well-Exposed1-After" border="0

Example 2 Before and After Grading:
00-Well-Exposed2-Before" border="0

00-Well-Exposed2-After" border="0

Example 3 Before and After Grading:
00-Well-Exposed3-Before" border="0

00-Well-Exposed3-After" border="0

These 3 examples yield between 62 and 64MB/s @2.5K 14bits, and the other clips I have top at 79MB/s which means that it is possible to get a guaranteed 30 seconds per clip (Max 1:45min with lvface, or 3:00+ with lvface+aewb hack) with a Sandisk Extreme 256, without having to heavily underexpose the image. With these parameters, the images are quite clean, with some light noise in the shadows, like any other camera, even the expensive ones... (See the previous video I posted: https://bit.ly/3eKo5e9)

Exposing a little hotter brings the noise lower, at the expense of shorter recording times. Which leads to the next test.

Examples recorded with Histogram showing 0.2

You can look at the Data Rates Here. Each group of 10-12-14 or 10-12-14-14 bits are the exact same image, recorded at different bit Depths. The extra 14 bit recording is to check consistency  or a lower exposure.
01-All-Files" border="0

Scene 1, very bright Highlights, 80.39 MB/s @2.5k 14bits
02-Scene-1" border="0

Scene 1 Graded
02b-Scene-1-Graded" border="0

Scene 2, Well balanced scene with some bright elements (The shiny objects are usually what chokes the recording). 74.27 MB/s @2.5k 14bits
03-Scene-2" border="0

Scene 2 Graded
03b-Scene-2-Graded" border="0

Scene 3, Well Balanced scene with less shiny spots (Allows a brighter overall image) 76.5 MB/s
@2.5k 14bits
04-Scene-3" border="0

Scene 3 with histogram at 1.0, lowers the data rate by 6.5 MB/s, and you get richer colors too
06-Scene-3-14bit-Hist-1-0" border="0

Scene 3 Graded
04b-Scene-3-Graded" border="0

Scene 4, Closeup with a lot of details and a very shiny object. This combination made the recording stop early at 14bits, with 91.23 MB/s
07-Scene-4" border="0

Scene 4 Graded
07b-Scene-4-Graded" border="0

Scene 5, Triggers frame corruption.

This scene, with a lot of details and reflective mosaic triggered a frame corruption error at 14 bits.
09-Scene-6" border="0

Graded
09b-Scene-6-Graded" border="0

Recorded with exposure right before it triggered a frame corruption, 91.59 MB/s at 2.5k, 14bits
12-Scene-6-Hist-0-2-right-before-corruption" border="0

Lowering Histogram to 1.0 and 2.0 yields a data rate of 88.1 and 76.16 MB/s respectively
10-Scene-6-Hist-1-0" border="0

11-Scene-6-Hist-2-0" border="0

When the image is that bright, bumping up the ISO has a negligible effect on data rates:
13-Scene-6-Hist-0-2-right-before-corruption-Iso-Var" border="0


Also note that moving the camera makes motion blur, and a blurry image decreases the data rate.
I won't post those clips, as this post is long enough already!

I did not let the camera record long in this test, to speed things up. For actual recording times, Have a look at the SD card tests I posted, which were in line with the Data Rates obtained here.

Maybe everyone could chime in as to what constitutes "a long enough clip recording time" to get the job done at higher resolutions.

Build Used: crop_rec_4k_mlv_snd_raw_only_2022Sep15.EOSM202
Cards: Sandisk Extreme Pro 128GB 170MB/s, Sandisk Extreme 256GB
Lens Ef-M 32mm F1.4

Danne

New build

Fixed pixelated bottom with 5.2k preset. Thanks masc for finding out.

vulcan4d

Got the SanDisk Extreme PRO 256GB 200/140 SDSQXCD-256G-GN6MA card , results were different in camera and playback modes @240Mhz

Camera mode: ~67MB/s write
Playback mode: ~95MB/s write

What do I believe? Is this normal?


cedricp

Is it possible to get a less cropped real-time liveview in 5.2K, or is it a limitation ?

Danne

Please read up on preview work done in the past. There's your answer.

theBilalFakhouri

Quote from: grusochsten on August 27, 2022, 08:14:02 PM
I'm having some difficulties with HDMI monitor to a Samsung Galaxy A52s phone on the latest build. I'm using a cheap capture card from Amazon.
...
...

HDMI and crop modes presets aren't well supported (at least in this build) and iirc they weren't fully tested with HDMI . .
I don't have EOS M, maybe Danne can help looking into these.

theBilalFakhouri

Quote from: roterabe on September 20, 2022, 08:49:34 AM
Isn't the new build referencing the 5D Mk.3? It could still be the right one, but I'm not that into the project to know what boot check determines on what camera I'm trying to run this.
...

Builds shared in this thread is only referring to EOS M.

theBilalFakhouri

Quote from: likwidchz on September 21, 2022, 03:03:45 AM
Tried it out, the 5k worked at ~1:23 seconds before the buffer filled up @ 14 bits.
Will have to try how long it does constant recordings, looks like @Gabriielangel mentioned 3 minutes. Did the camera auto stop or did you stop it?

If your SD card benchmarks ~90 MB/s write speed in Play mode:

There are many other factors for limited recording time, when using lossless compression it heavily depends on scene and ISO.
Other factors like if you are using hacks (under RAW video submenu) which boost write speed performance during RAW video recording.

Quote from: likwidchz on September 21, 2022, 03:03:45 AM
Btw, do people see any visual difference between 10-14 bits?

Yes, especially in shadows.

theBilalFakhouri

Quote from: julien becker on September 22, 2022, 01:23:39 PM
Danne and Bilal what you offer to the ML community is amazing ! Thank you so much.

Thanks, no problem :)

gabriielangel

Quote from: likwidchz on September 21, 2022, 03:03:45 AM
the 5k worked at ~1:23 seconds before the buffer filled up @ 14 bits.
Will have to try how long it does constant recordings, looks like @Gabriielangel mentioned 3 minutes. Did the camera auto stop or did you stop it?

On that particular test, the lvface+aewb hack was enabled,  and I stopped it myself at 3 minutes  (And I used the Sandisk Extreme Pro 170-128GB, with the regular sandisk extreme 256GB, I got about 1 minute ).
If you want to record longer, use 12bits, which will give you around 20 minutes continuous if your scene is not too bright. Like @theBilalFakhouri said, record time are scene and ISO dependent

gabriielangel

Hello, I shot outside today and I noticed something:

Issues with 5k frtp Preset
Build: crop_rec_4k_mlv_snd_raw_only_2022Sep29.EOSM202.zip

1
When pressing * or Half-Shutter to focus (focus-assist), the display becomes a lot brighter than the preview.
This makes it sometimes impossible to focus, as the display will turn completely white (Overexposed) if it is bright outdoors.

To replicate easily inside:

do not use a ND filter
Take a small lamp or lightbulb (I used a 60W Tungsten Bulb)
illuminate a small object or piece of paper
f3.5
ISO 100
Shutter 1/2200

This should be too dark for you to see anything in preview.
Press * or Half-shutter
You should now see the object clearly.

If you go outside in daylight, the display will get washed out when using focus-assist.

This does not happen when using the other presets.

2
When using focus-assist, it is not possible to navigate through the whole frame. The zone that is accessible is roughly a 4:3 square at the center of the display.
If there is something at the edges of the frame, it isn't possible to see it without panning the camera. This can lead to an out-of-focus image at large apertures (Because the distance changes, and the DOF is shallow)

Danne

New upload.

Skipping reduced framerate while previewing in x10 zoom. Tested before trying to resolve freezing bugs. Seems to do little so back to normal again.

2) Focus assist is a very good addition since it works while filiming. theBilalFakhouri built it and it´s pretty complex. Moving it around is not done easily.

EDIT:
Altered real time preview for Anamorphic 5.2k preset.

KnowledgeSupreme

Good evening. I am not sure if I am doing something wrong, but with the recent 5k update, the real time preview seems pixelated.

Danne


gabriielangel

Quote from: Danne on October 05, 2022, 07:21:12 AM
New upload.

Skipping reduced framerate while previewing in x10 zoom. Tested before trying to resolve freezing bugs. Seems to do little so back to normal again.

EDIT:
Altered real time preview for Anamorphic 5.2k preset.

Tested Today, 1 and 2 are fixed. Thanks!

A Good Way to Extend Recording Times  with High Dynamic Range Scenes at High Bitrates / Resolutions

I filmed outside in the sun today, and it was hard to get more than 10 seconds at 3K 1:1, because there were a lot of shiny surfaces in the frame. So I tried to lower the frame rate, and I were pleasantly surprised!

So, I took the opportunity to make a test with a moving swing. I am posting screenshots of the files, as I wouldn't have believed 16 fps could look that smooth.

With blowing tree leaves (Not included in the sample video), it showed a little as the movement was very fast, but at 20 fps, it was really good. As long as the Camera movements are deliberately slow, 20 fps is a good plan B when needed.

But as you will see in the example, the swing moves pretty fast...

Link to the Video: https://bit.ly/3SXUWeh


Files Screenshots:

MLV-Files" border="0

Da-Vinci-Screen" border="0

I weren't able to set the shutter lower (I guess I needed to tweak som regs), so I think the result would have been a little smoother at 16 fps.

The 5k Anamorphic flv preset is already at 20 fps, with the new high resolution. So those interested can try it right away.
Maybe you could bring back the 20fps mode for 3K (with the Enable 25fps trick)?


Danne

Enable shutter range full range to expand shutter expansion.
3k 24fps works here continuosly but I use a technique where I stay in cropped rel time preview and when checking for framing I enter framing shortly, then back to real time cropped preview and so on.

Might take a look at som fps reducing dial but don't hold you breath. Pretty busy over here

Danne

New upload:

- Added ability to increase/decrease fps through a submenu.
Will alter reg 6014 directly in the preset.
Note that changes will also be saved and has to be manually set to 0 again to return to original fps. It could be abused by selecting a too high fps which will stop the preset to start obviously. You then have to start the camera in photo mode and reset fps override from there.