1920x1080 RAW - real resolution?

[AndreyM]

Hi All,

Sorry if it's wrong place for my topic or if it's simple question. I got my 5D3 few days ago and started to test RAW recording in 1920x1080. Of course it looks way better than camera compressed videos but I'am still interresting if we get real 1920x1080 resolution. So I did few simple test coparing H.264, ML RAW and downscaled RAW photos. I may be wrong but my eyes over picture I prepared below saying that we have around 1600x900 as result of ML RAW.  Result is still great in comparison with H.264 onboard but...

Best Regards,
Andrey

Picture contain several 200% samples of H.264, ML RAW in 1920x1080, Full frame photo downscaled to 1600x900 and full frame photo downscaled to 1920x1080

[RenatoPhoto]

The only way to get the same image resolution as RAW is to use cropped digital video. 

How to record: -- RAW Digital Cropped Video on 5D3 --
http://www.magiclantern.fm/forum/index.php?topic=5441.0

This is a truly 1:1 sampling of the RAW sensor.  All other video modes use some line skipping.

[savale]

Good test! What would be nice to make the picture complete: can you make a raw video test 1600x900?

Downscaling a picture (either in photoshop or in camera) can be done using different algorithms. Each method is different and will have impact on the sharpness of the end result.

[dude]

Just to make it clear for me ( i did not understand everything):
If i choose like 1920x1080 for raw Video, it s downscaling like it does when using h.264.
And if i press magnification, for example 5x, it s not downscaled but takes this part from the sensor- right?
So is it just the resolution that s better or does it have more pro s on its side?

Thanks a lot!

[RenatoPhoto]

If i choose like 1920x1080 for raw Video, it s downscaling like it does when using h.264.

Yes but the image is in 14 bit res so it is way better that h.262 (8bit)

And if i press magnification, for example 5x, it s not downscaled but takes this part from the sensor- right?

Yes, it only samples a portion of the sensor and that is why we have to realign the focusing box with that portion of the sensor that is being sampled.

[dude]

Allright, got it.
But why is resolution bigger when in crop?

A lot of people do not understand that everything above 8 bit makes the files ways better then a better resolution. I don t understand this.
With more bith dept, you get rid of this fuckin dslr look and you have so much possibilities while grading.

[RenatoPhoto]

But why is resolution bigger when in crop?

My best guess:
I think Canon enabled this feature in LV so that you could zoom on the subject and do fine tune focusing.  Now ML captures that ability and turn it into DNG.

[Audionut]

Yes but the image is in 14 bit res so it is way better that h.262 (8bit)

1% also mentioned that there are various format changes in the Canon encoding to h.264, RAW - YUV - h.264 and that each one of these steps has a resize.

@AndreyM

There is clearly less detail and sharpness in the resized RAW images.  This could be caused by the resizing algorithm you used. 
And clearly less detail and sharpness in the RAW vs downsized photo.
Could you please update with images from a RAW x5 zoom (1:1 crop)?

[Shield]

I think for a more accurate test you'd have to use a resolution chart.  There were people saying the non-hacked MK3 did roughly 800 lines of resolution; the raw is definitely better.

Here's a thread on the subject:

http://www.dvxuser.com/V6/showthread.php?278877-5D-Mark-III-890-lines-of-resolution

For what it's worth, I've owned and shot with several GH2's and the AF-100.  The hack is right up there with those; perhaps a tad softer but with more dynamic range and better highlight control.  Still way better than a properly sharpened and color corrected H.264 though.

[Yoshiyuki Blade]

Can a video camera ever really resolve "true" 1080p? I recall there being a theorem (Nyquist something or other) that states that when converting things to digital, you need to acquire the source at twice the target resolution, then downscale it to maximize the quality at the desired resolution. This happens with things like flatbed scanners too. Scanning at 600 dpi looks worse than scanning at 1200 dpi and downscaling it to 600 dpi. Resizing algorithms can preserve the detail that was impossible to acquire at the native target resolution.

[mvejerslev]

I honestly think the RAW output at 1:1 is 'sharp enough', using the 5D Mark II..

[glnf]

Well, hello, hi! I'm new here - and, sorry, but this sort of discussions sometimes make me cringe a bit. So please allow me to add a bit of info...

What AndreM observes is correct. But the explanations given so far don't take into account the very basics of an RGB sensor. It is not possible to get the same resolution out of the sensor as when you scale a higher resolution image down. When shooting 1920x1080 we get this amount of pixels but they are divided into 1/4 Red 1/4 Blue 2/4 Green Pixels (in what is called a Bayer pattern). It would be save to assume that because of this we only get the actual resolution of 960x540 since we have to combine the different colors to create the color information. Good algorithms are capable of extracting a color image of a higher resolution but the resolution can't possibly reach the resolution of the sensor.

The only alternatives are 3 chip cameras where every pixel consists of 3 sub pixels, one for each color or a higher resolution chip downscaled. So a crop of the 5D3 sensor might be slightly better, since the camera doesn't have to apply the quick and rather rough down-scaling algorithm but it can't reach the the same quality as the "Full frame RAW photo downscaled" in AndreMs picture. You can find some more infos here: http://en.wikipedia.org/wiki/Bayer_filter. Said that, to me the quality achievable with a RAW 5D3 is absolutely astonishing and brings the camera in a league with high-end cine-cameras that cost a fortune. So nothing to worry about really.

[KMikhail]

glnf is correct.

Bayer pattern has about 80% effective resolution for B/W (like charts) and 50% for R/B. G is in the middle. This is what we see.

Full frame video RAW looks a bit worse than a 1:1 due to the facts

a) it likely uses primitive binning
b) AA filter atop of the sensor is too weak for the 1920x1280 over the 36x24mm, so it doesn't help with moire removal.

[IliasG]

Well said, by glnf and KMihail,

But there is one parameter missing. It is the nature of the sampling in a grid that firstly decreases resolution and the Bayer filter adds a bit more decreasing. And the aa filter a bit more.

For a consistently proper sampling of a line pair with a monochrome grid sensor we need at least 3 rows or sensels or else we depend on the (by coincidence) aligning of the detail with the grid or not. This varies from 100% with perfect aligning to 0% when it's totally off. And on average for a reconstruction with MTF-50 the resolution is around 0.7X of the grid lines. This 0.7X sounds way small because our standard when inspecting resolution is Lower than MTF50 .. MTF30 perhaps ..

The Bayer filter and the Debayering process decrease the result a bit more .. it's efficiency vs the monochrome is around 90% .. http://bvdwolf.home.xs4all.nl/main/foto/bayer/bayer_cfa.htm.

Saying that the sampling density of the 3 RGB colors on a bayer CFA is equal to the distance between the same channel sensels is not accurate because the bayer R,G1,B,G2 are not the same as the RGB R, G, B pixels. The "bayer Red" sensel samples both R & G (and a bit of B) colors of RGB color space and so do the rest. This info used by clever demosaic algos gives a reconstruction with 90% efficiency vs the monochrome.

[AndreyM]

@AndreyM

There is clearly less detail and sharpness in the resized RAW images.  This could be caused by the resizing algorithm you used. 
And clearly less detail and sharpness in the RAW vs downsized photo.
Could you please update with images from a RAW x5 zoom (1:1 crop)?

I did several video RAW shots with 1:1 crop and it's appears excatly the same quality in comparison with Full Frame photo in RAW.
By the way I love this 1:1 crop and it was the reason I was still using 600D (since it is the only Canon DSLR providing this mode natively). For example full moon rise here http://www.youtube.com/watch?v=NP7hy5R9MLs I shot with Canon 600D + EF 70-200 lens.
Now having this feature on 5D3 with ML in 14 bit RAW is awesome !!!

Thanks Magic Lantern and everybody contributing!

Good test! What would be nice to make the picture complete: can you make a raw video test 1600x900?

I assume if shot in 1600x900 RAW and upscale to 1920x1080 it will have about the same visible resolution (or even litle bit better) as 1920x1080 camera H.264.

Best Regards,
Andrey

[aaphotog]

What is this 1:1 crop thing?

is it still a 16x9 picture?
Does it give the same FOV as the normal 1920x1080 Raw recording?

[RenatoPhoto]

The only way to get the same image resolution as RAW is to use cropped digital video. 

How to record: -- RAW Digital Cropped Video on 5D3 --
http://www.magiclantern.fm/forum/index.php?topic=5441.0

This is a truly 1:1 sampling of the RAW sensor.  All other video modes use some line skipping.

Please read before you post!  Many aspect ratios available and no, it is not same FOV.

[AndreyM]

Another test comparing frame from 1:1 crop RAW at 24mm (upper image) and normal video RAW at around 72mm(lower image). Visible resolution is about the same while 1:1 have more visible noise, I shot with ISO 1600 though.

Best Regards,
Andrey