Some questions about specific terms around ML

Started by Calussien, August 05, 2020, 05:25:14 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Calussien

Hi ML users,

I'm new to the ML world, I just discovered this forum and read almost every topic around 5DM3/M2 Raw capabilities and other stuff.
I am a young french camera assistant that just begun working in the cinema industry and I'm super interested in understanding all of the video worflow from light captured by the sensor to the post-production and deliver to customers (Cinema/TV/internet).

Through my researchs, I've been faced with a lot of technical terms that seems really important if I want to get the best from ML or simply to understand how video camera works. Can somebody help me answering some of my questions ? I think this post could be very useful for new user of ML like me.

Pixel binning : How does it work with a CMOS RGGB sensor ? Is it a process done before or after photocytes charges are read and turned into data ? Does it really helps reducing noise ?
Also, I've seen a 5.7k anamorphic build for the 5DM3 that use 1x3 bining 1920pix then unsqueeze the raw file in post to get back the 5760 horizontal pixel from the sensor. Does it imply pixel binning can be reversed in post ?

1:1 pixel readout mode : So, it means that all photocytes in the captured area of the sensor are recorded ?

Oversampling and downsampling : I'm really confused about the difference about those two (if there is any) and when can we use them. I've seen a debayer process that combines 2x2 pixels to form "super" RGB pixels. Is this downsampling ? Can we downsample an RGB or YCbCr picture already debayered ? In a 1080p screen (like my laptop), a UHD video file downsample to 1080p and the same UHD video exported full res will appear different ?

lossless RAW : I guess it's a RAW video file that is compressed with a lossless compression ? So, different than uncompressed RAW

Debayer process : This is a huge one. I know what's the idea behind it and already found some articles about things like bilinear or AMaze. I'm more interested about the reasons to choose a process instead of another in MVL App. And, in Davinci Resolve, I never saw an option to choose a specific process. So do you know what debayer process software uses ? Does it depend on the camera that shoot the file ?

SD overclocking and card spanning : Can you explain what are they ?

Upscaling : 5DM2 can record 1824x1026p RAW video. If I want to deliver a 1080p video, I need to upscale my file. How does it work with pixels ?

Centered crop mode : What is the meaning of "centered" ? Does it mean other crop modes aren't centered ?


Thanks a lot and have a nice day.

Calussien

12georgiadis

+1 for 1X3 it's still not clear for me, especially for 5D3 5.7K mode

muanthang

Quote from: Calussien on August 05, 2020, 05:25:14 PM
Hi ML users,

I'm new to the ML world, I just discovered this forum and read almost every topic around 5DM3/M2 Raw capabilities and other stuff.
I am a young french camera assistant that just begun working in the cinema industry and I'm super interested in understanding all of the video worflow from light captured by the sensor to the post-production and deliver to customers (Cinema/TV/internet).

Through my researchs, I've been faced with a lot of technical terms that seems really important if I want to get the best from ML or simply to understand how video camera works. Can somebody help me answering some of my questions ? I think this post could be very useful for new user of ML like me.

Pixel binning : How does it work with a CMOS RGGB sensor ? Is it a process done before or after photocytes charges are read and turned into data ? Does it really helps reducing noise ?
Also, I've seen a 5.7k anamorphic build for the 5DM3 that use 1x3 bining 1920pix then unsqueeze the raw file in post to get back the 5760 horizontal pixel from the sensor. Does it imply pixel binning can be reversed in post ?

1:1 pixel readout mode : So, it means that all photocytes in the captured area of the sensor are recorded ?

Oversampling and downsampling : I'm really confused about the difference about those two (if there is any) and when can we use them. I've seen a debayer process that combines 2x2 pixels to form "super" RGB pixels. Is this downsampling ? Can we downsample an RGB or YCbCr picture already debayered ? In a 1080p screen (like my laptop), a UHD video file downsample to 1080p and the same UHD video exported full res will appear different ?

lossless RAW : I guess it's a RAW video file that is compressed with a lossless compression ? So, different than uncompressed RAW

Debayer process : This is a huge one. I know what's the idea behind it and already found some articles about things like bilinear or AMaze. I'm more interested about the reasons to choose a process instead of another in MVL App. And, in Davinci Resolve, I never saw an option to choose a specific process. So do you know what debayer process software uses ? Does it depend on the camera that shoot the file ?

SD overclocking and card spanning : Can you explain what are they ?

Upscaling : 5DM2 can record 1824x1026p RAW video. If I want to deliver a 1080p video, I need to upscale my file. How does it work with pixels ?

Centered crop mode : What is the meaning of "centered" ? Does it mean other crop modes aren't centered ?


Thanks a lot and have a nice day.

Calussien
"There are some questions that shouldn't be asked until a person is mature enough to appreciate the answers."
― Anne Bishop, Daughter of the Blood

theBilalFakhouri

Quote from: Calussien on August 05, 2020, 05:25:14 PM
Pixel binning : How does it work with a CMOS RGGB sensor ? Is it a process done before or after photocytes charges are read and turned into data ? Does it really helps reducing noise ?
Also, I've seen a 5.7k anamorphic build for the 5DM3 that use 1x3 bining 1920pix then unsqueeze the raw file in post to get back the 5760 horizontal pixel from the sensor. Does it imply pixel binning can be reversed in post ?
https://www.magiclantern.fm/forum/index.php?topic=16516.0
Binning is done in analog domain, We can't Un-bin the pixels is post, we unsqueeze RAW file to correct the aspect ratio, so there is a quality loss if you compared it to 1:1 (Read every pixel on sensor without Binning or Skipping), Example: Look how the details between 1x1, 1x3 , 3x3:
https://www.magiclantern.fm/forum/index.php?topic=16516.msg210023#msg210023
Ignore 3x1 image, it's have jagged line issue, but re-solved after I made the tests, didn't make a new test showing true quality of 3x1.

Yes, Binning does reduce the noise.

Quote from: Calussien on August 05, 2020, 05:25:14 PM
1:1 pixel readout mode : So, it means that all photocytes in the captured area of the sensor are recorded ?
All pixels are been recorded, without Binning or Skipping


Quote from: Calussien on August 05, 2020, 05:25:14 PM
Oversampling and downsampling : I'm really confused about the difference about those two (if there is any) and when can we use them. I've seen a debayer process that combines 2x2 pixels to form "super" RGB pixels. Is this downsampling ? Can we downsample an RGB or YCbCr picture already debayered ? In a 1080p screen (like my laptop), a UHD video file downsample to 1080p and the same UHD video exported full res will appear different ?
https://www.youtube.com/watch?v=AcO0S1An_LY


Quote from: Calussien on August 05, 2020, 05:25:14 PM
lossless RAW : I guess it's a RAW video file that is compressed with a lossless compression ? So, different than uncompressed RAW
Lossless RAW use lossless compression, you can decompress it using e.g MLVApp to get the full size of uncompressed RAW again, no quality difference, no quality loss, at all.


Quote from: Calussien on August 05, 2020, 05:25:14 PM
Debayer process : This is a huge one. I know what's the idea behind it and already found some articles about things like bilinear or AMaze. I'm more interested about the reasons to choose a process instead of another in MVL App. And, in Davinci Resolve, I never saw an option to choose a specific process. So do you know what debayer process software uses ? Does it depend on the camera that shoot the file ?
Google has good answers, MLVApp uses open source de-bayers, some are better than others in some aspects, Davinci Resolve might use it's own developed de-bayers, and they simply didn't include more than one.

Quote from: Calussien on August 05, 2020, 05:25:14 PM
SD overclocking and card spanning : Can you explain what are they ?
Canon has limited SD card controllers to 40 MB/s in most cameras that use ML, however the controllers support up to ~104 MB/s, SD overclocking unlocks the write/read limit from Canon, not all cameras are stable in these high write/read speeds.

Card Spanning, is to use both CF and SD card for recording, in ML it use it to record high resolution RAW video that requires around more or less 120 MB/s

CF card can reach up to ~85 MB/s write speed on 5D3
SD card can reach up to ~60 MB/s write speed on 5D3 (Maybe more)
Card spanning will combine the two write speeds from the CF and SD, to get up to 130 MB/s in Video mode.

Quote from: Calussien on August 05, 2020, 05:25:14 PM
Upscaling : 5DM2 can record 1824x1026p RAW video. If I want to deliver a 1080p video, I need to upscale my file. How does it work with pixels ?
?

Canon upscale it to 1080p while doing H.264 Encoding, but the RAW data is 1824x1026p on 5D2 , there is no native 1080p in non-crop mode, except for 5D3.

Quote from: Calussien on August 05, 2020, 05:25:14 PM
Centered crop mode : What is the meaning of "centered" ? Does it mean other crop modes aren't centered ?

All presets I think it's centered, but this one will not modify the resolution in x5 which is 3584x1320 in 5D3, by using centered crop mode it will center the recording area of sensor for you if it's not centered already.

These are a quick answers, it may clear up some things for you.


reddeercity

Quote from: Calussien on August 05, 2020, 05:25:14 PM
...... 5DM2 can record 1824x1026p RAW video. If I want to deliver a 1080p video .........
No 5D2 records 10bit 1880x1248 @24 or 30p , for 16x9 or 1.778 A.R. 1880x1056 .
Simply export to 1080p & the software will scale it correctly , no different then the Pro $50K camera .
Along as you go from raw->intermediate codec->consumer device friendly codec (h264/h265) , I scale at the intermediate codec level(prores 4444xq)
but you can scale at the final export h264 codec if you like , I fine sooner in the workflow results in a better scale to 1080p .
Or got can go from raw-> final export , depending on your software .

Calussien

Many thanks to all of you. Everything makes a bit more sense now and I think it may help others too.  :)

I'm still curious about this debayer process called "Superpixel" which looks like some sort of downscaling and debayering combined.

Look at this post : https://pixinsight.com/doc/tools/Debayer/Debayer.html


12georgiadis

Quote from: theBilalFakhouri on August 12, 2020, 10:32:51 PM
https://www.magiclantern.fm/forum/index.php?topic=16516.0
Binning is done in analog domain, We can't Un-bin the pixels is post, we unsqueeze RAW file to correct the aspect ratio, so there is a quality loss if you compared it to 1:1 (Read every pixel on sensor without Binning or Skipping), Example: Look how the details between 1x1, 1x3 , 3x3:
https://www.magiclantern.fm/forum/index.php?topic=16516.msg210023#msg210023
Ignore 3x1 image, it's have jagged line issue, but re-solved after I made the tests, didn't make a new test showing true quality of 3x1.


It would be interesting to compare 5.7K 1X3 with Topaz video enhancer that upscales 1:1 to 5.7K as well

masc

Quote from: 12georgiadis on August 13, 2020, 01:13:30 PM
It would be interesting to compare 5.7K 1X3 with Topaz video enhancer that upscales 1:1 to 5.7K as well
The biggest difference is the depth of field, and this can't be simulated by any software. Here just the used physical size of the sensor counts. And this size is just ~ the half for 1x1 reading, while the read pixel count is not sooo different. So 1x1 is just slightly sharper, while 1x3 looks way more cinematic (for my taste).
5D3.113 | EOSM.202