Bilal's crop_rec_4k experiments for 650D / 700D (T4i / T5i)

[ML700D]

thanks bilal, I just did and works fine.

[disorderr]

Hi, i've been struggling with the shutter blanking isoless build, everytime i stop recording the timer freezes and ML menu shows "Stopping" but it doesn't stop.
Any idea of what can it be?
BTW, thanks bilal, amazing work, i love your builds <3

[lightwriter]

Hi Bilal!

I've been watching some videos from Zeek, and I found this one quite interesting:

In minute 2:16, it says that the EOS M 2.5K RAW mode is ~3.2x Crop. This poses some questions:

- Does this mean that it has double the magnification of the normal 1.6 crop from Canon, i.e., does it use half the sensor area? If that's so, maybe it's not the preferred method when filming in low light, as it gathers less light, am I right?

- Is 2.5K Expanded Preview similar to EOS M 2.5K RAW mode?

- What's the setting that allows to record with the most resolution possible, with no moiré, small crop factor and real-time preview? 4.5K 1x3, 4.3K 1x3, 4K 1x3 and UHD 1.3 all offer about 1 megapixel of resolution. Is there something I can do to increase it to 1,5 or even 2 megapixel resolution? I am already using the SD hack at 240Mhz.

[theBilalFakhouri]

Hello!

In minute 2:16, it says that the EOS M 2.5K RAW mode is ~3.2x Crop. This poses some questions:

- Does this mean that it has double the magnification of the normal 1.6 crop from Canon, i.e., does it use half the sensor area? If that's so, maybe it's not the preferred method when filming in low light, as it gathers less light, am I right?

Yes it's doubled magnification compared to 1.6 crop, it uses ~1/4 of the sensor area:

(16:9 showcase)

- Is 2.5K Expanded Preview similar to EOS M 2.5K RAW mode?

"2.5K Expanded Preview" is same as "2.5K 1:1 centered frtp", However *"1440p" on 700D is same as  "EOS M 2.5K RAW".
*"1440p" preset on 700D has 2560x1440 resolution which can be called 2.5K. while "EOS M 2.5K RAW" has 2520x1418 resolution (a little less than 700D).

To get 2.8K on 700D, you can use "3K 1:1" preset which has 3072x1308 resolution, reduce the resolution to 2880x1226 via RAW video submenu (mlv_lite), now it can be called 2.8K .

- What's the setting that allows to record with the most resolution possible, with no moiré, small crop factor and real-time preview? 4.5K 1x3, 4.3K 1x3, 4K 1x3 and UHD 1.3 all offer about 1 megapixel of resolution. Is there something I can do to increase it to 1,5 or even 2 megapixel resolution? I am already using the SD hack at 240Mhz.

All of 1x3 presets with real-time preview has ~2.7 Megapixel (out of the sensor, of course in 1x3 readout), not 1 MP.
Use one of 1x3 presets depending on aspect ratio you need.

Downscaling e.g. 1280x2160 to 1280x720 will give cleaner image, upscaling 1280x2160 to 3840x2160 will give more details.

You can go above 2.7 MP using "Crop mode V2" (crop_new) on 700D, but you will lose real-time preview. The reason why all of 1x3 real-time presets are limited to ~2.7 MP because of preview, going above ~2.7 MP resolution will make Preview frozen or you will lose part of the horizontal /vertical preview (there is a limit here, more research is required).

Maximum resolution in 1x3 Binning mode @ 23.976 FPS with full width sensor is 1736x2214 2.35:1 AR which is 3.8 MP, this mode isn't continuous. But we can achieve 1600x2040 2.35:1 AR 1.74 crop factor @ 23.976 FPS in 1x3 and have continuous recording (10-bit lossless, 240 MHz SD overclock), that's about 3.2 MP, but without real-time preview (using "Crop mode V2", set resolution to 1736x2214 then reduce it with RAW video to 1600x2040).

[ML700D]

btw bilal,
why when use 1fps overrides lv looks very choppy, lag and changing darker then brighter like that?

[lightwriter]

Thanks Bilal, for your very thourough answer!

You're right, the 1x3 presets have ~2.7 Megapixel; however, since the vertical resolution must be squezzed to a third, the final resolution results in 1 Megapixel. 1280x2160 (using your example) gives us a final result of 1280x720 = 921600 pixels, since 2160 must be divided by 3, correct?

I tried the 1600x2040 2.35:1 AR 1.74 crop factor @ 23.976 FPS in 1x3 and used the EDMACs hack. When I was recording I had live preview, but it was distorted (stretched vertically). Is this what I should expect?

This results in a very good quality footage; however, in the MLV App a horizontal stripe artifact showed at the bottom of the image. Do you have any idea why did this happen?

Bottom artifact:

[theBilalFakhouri]

Thanks Bilal, for your very thourough answer!

.. since the vertical resolution must be squezzed to a third, the final resolution results in 1 Megapixel. 1280x2160 (using your example) gives us a final result of 1280x720 = 921600 pixels, since 2160 must be divided by 3, correct?

You are downscaling instead of upscaling, this result in less details and resolution, this way you are getting oversampled 1280x720. Instead of dividing 2160 by 3, multiply 1280 by 3 which result in 3840x2160 1x3 image which has more details compared to 1280x720 (of course it's not true 3840x2160, also it's not 1280x720, it's actually in-between these resolutions. I call it myself Hybrid-1080p).

MLVApp upscale 1x3 footage by default by multiplying horizontal resolution by 3. Transformation tab set to "0.33" when importing 1x3 footage.

I tried the 1600x2040 2.35:1 AR 1.74 crop factor @ 23.976 FPS in 1x3 and used the EMACs hack. When I was recording I had live preview, but it was distorted (stretched vertically). Is this what I should expect?

This results in a very good quality footage; however, in the MLV App a horizontal stripe artifact showed at the bottom of the image. Do you have any idea why did this happen?

Bottom artifact:

You need to refresh LiveView manually (after changing Crop mode V2 settings) by pressing SET button in LiveView (before recording).
Switch Preview under RAW video submenu to Framing (or press half-shutter in LiveView) to check if these artifacts gone after applying a manual refresh.

[lightwriter]

You are downscaling instead of upscaling, this result in less details and resolution, this way you are getting oversampled 1280x720. Instead of dividing 2160 by 3, multiply 1280 by 3 which result in 3840x2160 1x3 image which has more details compared to 1280x720 (of course it's not true 3840x2160, also it's not 1280x720, it's actually in-between these resolutions. I call it myself Hybrid-1080p).

MLVApp upscale 1x3 footage by default by multiplying horizontal resolution by 3. Transformation tab set to "0.33" when importing 1x3 footage.

Thanks, I didn't know that! I filmed a bottle of water with very fine details in the mode UHD 1x3 - 1280x2160. Then, in MLV App I extracted a frame at 1280x720 and another at 3840x2160. Afterwards, I downscaled the 3840x2160 to 1920x1080 and upscaled the 1280x720 one to 1920x1080, as you can see:



I did the same process again, only this time downscaling the 3840x2160 to 2560x1440:



To my eyes, there's definitely an improvement in the first case. The fine details in the paper, especially the green stripes, are visible when downscaling to 1920x1080, being invisible when the image is exported at 1280x720. You said the 1280x720 would have less detail but would be cleaner, as it can be seen. However, I think is worth it to uspcale and denoise it; the software that I use (RG Magic Bullet) can clean the noise and maintain the vast majority of the detail.

When downscaling to 2560x1440 I don't think there's extra information in relation to 1920x1080. What do you think is the "real" resolution when exporting the footage at 3840x2160?


Relative to the second question, your step-by-step guide helped me get rid of the corrupted stripe, thanks!

[ML700D]

nice result.. 👍
UHD 1x3 is my favorite..
but bilal's crop_rec_4k is the real favorite.. 🎥

[IDA_ML]

Absolutely!

The UHD 1x3 1280x2160 mode at 12bit and 24 fps provides an absolutely gorgeous result.  Recording is continuous too.  I use it all the time.

For wide  angle landscape videography, there is another favorite mode of mine: the 5k 1x3 1736x2928 mode at 12bit and 16 fps.  This mode is also continuous and provides full-sensor readout.  Works great with Dual ISO too - perfect for sun sets or other high-contrast scenes.  If you set the shutter to 1/16-th of a sec. you also get an additional stop of light - perfect for street videography at night.  You get a nice motion blur but need to keep the camera very steady (tripod, gimbal), otherwise rolling shutter will kill your shots.

All in all I think that the 1x3 high-resolution modes are among the greatest achievements of our developers.

[lightwriter]

Thanks ML700D!

Indeed IDA_ML, it's amazing the quality the developers achieved with the 1x3 high-resolution modes with so ancient cameras...

[theBilalFakhouri]

New builds are posted in the first post for 650D/700D

Changes:

mlv_lite:
-Added new LiveView hacks
-Get up to ~13 MB/s write speed improvement! total write speed in LiveView is 82.8 MB/s! :

Guide: LiveView hacks (write speed improvement)

-Some results:
-2560x1440 1:1 16:9 Continuous! (@ 23.976, 10-bit lossless)
-1736x2214 1x3 2.35:1 Continuous!  (@ 23.976, 10-bit lossless)
-2880x1226 1:1 2.35:1  Continuous!  (@ 23.976, 10-bit lossless)
-You can definitely record longer in 3K 3072x1308 1:1 2.35:1 preset (@ 23.976, 10-bit lossless)
-High frame rate modes can be Continuous! or "You can definitely record longer".

-Old EDMACs hack has been removed and replaced with the new hacks.

crop_rec:
-Re-enabled some ADTG registers.

[ML700D]

nice.. thanks bilal 🥳

[theBilalFakhouri]

btw bilal,
why when use 1fps overrides lv looks very choppy, lag and changing darker then brighter like that?

Sorry for the delay,

LiveView should look choppy because it's low FPS.
Laggy/the camera become slower: overriding FPS to low frame rates changes FPS Timers (LiveView speed is based on FPS timers) which result in slower LiveView updates.

Changing darker then brighter: This is related to shutter blanking and how it work, it will takes more timer to restore and apply the correct shutter speed. adjusting FPS to lower framerate will result in limited and slow shutter speed --> bright LiveView --> Shutter blanking will take aware of this and will restore full shutter speed range and will apply previous shutter setting --> dark LiveView (actual brightness before changing FPS).

[theBilalFakhouri]

What do you think is the "real" resolution when exporting the footage at 3840x2160?

I am not sure if there is answer for that . . however 1280x2160 is ~2.7 MP, and the upscaled footage is based on that number of pixels.
Also the 1280 width resolution is based on 3840 Binned pixels from the sensor.

[lightwriter]

Hi Bilal!

Yes, I can believe it's hard, if not impossible to reach to a number.
However, in your opinion, do you believe it's equivalent to 1920x1080, or a lower resolution?

[names_are_hard]

What do you mean by "real"?  1280*2160 should have all the light from 3840*2160 because it's binned, and at least as good resolution for detail as 1280*2160 (probably, slightly better angular resolution?  Binning loses you some but you don't need as good a lens etc to have the resolution available over the 3 pixels as you would over 1?).  If you downsample to below 2160 height, you are throwing away resolution.  The *vertical* resolution really is 2x 1080.

So, it should definitely be better than 1920x1080 resolution overall.  That's 2.1MP.  1280*2160 is 2.7MP.  But they're a little odd to compare.  The horizontal and vertical angular resolution aren't the same in 1x3 binned, but the native 1920x1080 is binned anyway and compressed so what are you comparing against?  1920x1080 raw from some different camera?

[lightwriter]

What do you mean by "real"?  1280*2160 should have all the light from 3840*2160 because it's binned, and at least as good resolution for detail as 1280*2160 (probably, slightly better angular resolution?  Binning loses you some but you don't need as good a lens etc to have the resolution available over the 3 pixels as you would over 1?).  If you downsample to below 2160 height, you are throwing away resolution.  The *vertical* resolution really is 2x 1080.

So, it should definitely be better than 1920x1080 resolution overall.  That's 2.1MP.  1280*2160 is 2.7MP.  But they're a little odd to compare.  The horizontal and vertical angular resolution aren't the same in 1x3 binned, but the native 1920x1080 is binned anyway and compressed so what are you comparing against?  1920x1080 raw from some different camera?

Yes, that's it, I'm trying to compare it against 1920x1080 raw from a different camera. If the vertical resolution is really 2x1080, what happens to the horizontal resolution? Are the 1280 pixels "stretched", i.e. interpolated, if we want to maintain the 2160 vertical dimension and 16:9 ratio? Sorry if my question is dumb, I understand this involves a lot of math.

[Walter Schulz]

Why don't you take some shots with test charts? Newton rings, Siemens stars, ...

[names_are_hard]

Right, so you're comparing two different cameras, two different optical paths.  Using only the resolution numbers isn't going to get you a good answer.  As Walter says, take some calibrated test shots.

Yes, the vertical resolution is real 2160.  Horizontally, it's binned pixels; the light from every 3 pixels is combined in hardware.  Light is captured from a 3840*2160 region, the data produced is 1280*2160, but the pixels are rectangular.  The "stretching" is to undo the assumption by image formats that pixels are square.  Stretched and interpolated are not the same thing.  I believe this is not interpolation, but I haven't checked the code.  There is very little maths involved here.

It is definitely higher resolution than 1920*1080 *in terms of pixels*.  This doesn't tell you if it looks better or worse.  If you want to know that, you must measure it.  There's enough of an MP difference that I'd expect the ML footage to look noticeably better, but that's my guess.

1920*1080 is also a fairly weird raw format.  What cam is it?

[Skinny]

1920x1080 doesn't leave any margin for stabilization or cropping, 3840 even stretched from 1280 but with real 2160 vertical should be much better in this aspect.. at least I think so. Of course I can only compare it to ML 1920x1080 and not other cameras.

[mlrocks]

I was trying to use the silent picture function in video mode on 650d.
I noticed that when coupling a dumb adapter with a manual focus lens, push the shutter button will generate error like "communication issue with lens". Then I have to remove the battery to clear.
Today I try connecting Canon EFS 10-18 IS, then push the shutter button, the camera generates both CR and JPEG files, and I hear a shutter sound. I have the silent picture on with "simple dng". I am happy that I can take pictures in video mode, although silent picture function seems not working, which does not bother me much.
It seems to me that the lens connection plays an important role when trying to take photos in video mode. Hope this can solve so that manual focus lens with dumb adapter will take photos too.

Update:
I just found out that the error was caused by my using extender. If there is no extender, open and close the battery, restart the camera, mf lenses with dumb adapters can take photos, although silent picture seems still not work, but this is good enough for me. 

[theBilalFakhouri]

New builds are posted in the first post for 650D/700D

Changes:

mlv_lite:
-Excluded "Movie crop mode" from the new hacks, it freezes LiveView.
-Added a little more delay after calling CartridgeCancel a.k.a "One more hack", might help reducing/fixing first frame corruption.

-Menus:
-fixed interference between ML menu and Canon overlays in LiveView, I took this fix from here, solves this issue.

[lightwriter]

Hi Bilal!

Thanks for the fixed interference between ML menu and Canon overlays in LiveView. At times it could be very annoying.
Keep up the good work!

[lightwriter]

Hi again Bilal Fakhouri!

Is it somehow possible to shoot at 30 fps in the UHD 1x3, 4K 1x3 or 4.3K 1x3 modes?
Or if I set some values manually, can the 700D with all the speed hacks actually achieve this?