Exclusive: here is the first app to resample Panasonic GH4 4K 8bit to 10bit 4:4:4

[mtheory]

I am really happy with the GH4 and A7s developments, and hope Panasonic/Sony will now provoke Canon into a new DSLR industry Cold War. Canon reps looked miserable at all the NAB interviews. They have no choice but to give the next 7D2 a 4K video resolution now...just imagine the ML RAW on a superior hardware!

[skiphunt]

What glass was used on these clips? Compared to the Sony A7S stuff, this converted GH4 footage looks soft to me.

[see ya]

Really? Misleading? Read the damn article Yellow.

 

I did read the previous articles, my comments are all there to see, what I find misleading is that and this is not meant as heavy criticism just address the balance, of coarse this is your site, you tell it how you want, "read it here first, exclusive" is an old adage and sells newspapers, but also the forum is here to provoke discussion rather than merely ass kiss?

 

The articles center on the GH4, fine, talking it up may help prevent it becoming marginalized in the slew of other first round of rec709 shooting 4K cameras from other manufacturers. But the questionable process of 4:2:0 to 4:4:4 8 to 10bit can be done on any camera source, any reasonable resolution although little point in 1080p down sampling. The Canon 1DC 4k, the Nikon V1 4k ??

 

The articles suggest 4:4:4, it wasn't until David Newman clarified he considered that it was RGB 4:4:4 not YCC 4:4:4, big difference and to many it was read as YCC, big news 4:4:4 YCC, well no not really.

 

And 4:4:4 RGB is a particular description to suggest natively captured full sample RGB, not RGB from a 4:2:0 compressed via some interpolation and scaling down scheme, even with the theoretical maths.

 

Then 10bit output, well actually no, 10bit in the luma, again dubious as to the extent of the benefit of that, comments that it grades better, it's not any surprise that scaling down and interpolating values makes it 'better' for grading, that's been common knowledge for a long time, in fact taking an 8bit 4:2:0 frame, converting to RGB using bicubic or Lanzos interpolation rather than point resize provides cleaner edges, interpolated values to mush the image up a bit and therefore appear to grade better than a 4:2:0 YCC frame, simply applying a tiny blur to a 4:2:0 frame interpolates the pixels and it stands a bit more grading, denoising a 4:2:0 frame at 32bit say, will give interpolated values and make it appear to grade better. BUT bottom line is all that can be done in the NLE or grading app, not some preprocess transcode, eating mass storage and 'better' for grading what compared to what?.

 

Does the image actually look any better compared to say a 32bit workflow, just how much of the benefit here is from the 10bit aspect or even the pseudo 4:4:4 and how much is from the scaling down and interpolation of values in the conversion to RGB?

 

But every ones own tests will decide that for themselves.

 

What's getting shouted about is the 10bit and the 4:4:4 like it's gold standard output from the GH4 albeit at lower resolution, but as emphasized now 2k, that's still serious, talking the camera up in the process like it's something special other 4k cameras can't do.

 

On the new discovery, I feel it's misleading to suggest that this new found process of alchemy happened here via this site, yes Thomas has provided the app as a first but it's not too far stretched to consider that the process has been done before, anyone using Matlab, Avisynth or even Nuke or Shake has probably done the math. 4K to 2K, it's common to go 4:4:4 RGB 4K film scan native from the scanner sensor or from a 4:4:4 RGB camera like the Alexa rather than YCbCr 4:2:0 compressed using h264 which not only throws away by subsampling but also throws away by compression. It's not 4:2:0 uncompressed nor 4:2:0 with gentle compression.

 

So this process works any resolution, any 4:2:0 or 4:2:2 camera source, not a new process, RGB 4:4:4 interpolated, not native. Mileage will vary.

[Baremis]

I have a question : If I shoot in 4k the sensor 4/3 crops 2x ? is it smaller than super 16mm gate size ? 

[Joseph Moore]

M43 is bigger than super16, but yields a similar feel, to me. With the use of a focal reducer, you can get a FOV more similar to Super35

[Joseph Moore]

I don't think anyone claimed that this was a new process. What is new is a very low priced tool for acquiring an image sufficiently high resolution to use this technique to generate a 2k image.

 

Supersampling a 4k image and interpolating to 2k yields true 4:4:4 RGB with more real color information (and a lot more luma information) than if shot at output resolution at 4:2:0. That math is pretty easy to understand.

 

All the nitpicking aside, the simple take away is that there are compelling reasons to acquire in 4k even if your delivery medium is 2k/1080p.

[see ya]

Supersampling a 4k image and interpolating to 2k yields true 4:4:4 RGB with more real color information (and a lot more luma information) than if shot at output resolution at 4:2:0. That math is pretty easy to understand.

 

My comments have nothing to do with dissing 4k to 2k, cheaper 4k cameras, compelling reasons to aquire 4k or anything else you care to mention.

 

My comments relate to the dubious suggestion that "true 4:4:4 RGB" is created from a 4:2:0 h264 source however easy you think the maths. If you don't mind me asking what's your understanding of what "true 4:4:4 RGB" is compared to other methods of constructing RGB frames from 4:2:0 h264?

[krisicwalsh]

Just a thought, would it be possible to resample e.g. 96fps footage to 4:4:4 through some similar process and see an appreciable difference?

 

Great job Thomas with the app, it certainly does seem like a nice tool especially as Andrew mentioned for B&W.  Maybe we don't all need to go out and buy the new Digital Bolex to get shots of our leaky taps, after all. (Kidding, love what I've seen of the 12 bit raw on the D16M; this is super versatile, the Bolex is a special application camera.)

 

Christopher Saalbach-Walsh

@krisicwalsh

[leeys]

Technical arguments aside... how well does it work practically? I'd like to know compared to a 2k 8 bit compressed file and a 2k raw file, how well does a 4k supersampled to 2k file grade? Will it be closer to the 8 bit file than the raw file?

[Kolen Cheung]

The Math is easy to explain.

 

First, 4:2:0 means it has half of the resolution in both directions in the chroma channel.

So if we downsample the resolution to a half, we have full color information in both directions, hence 4:4:4.

 

Second, originally the luminance has full resolution and are 8 bit. We can consider 8 bit precision means there are 256 possible values, and as each of them are precise up to an integer, the absolute error is 0.5. Averaging 4 pixels will divide the absolute error by 4, hence 1/8. To make it an integer, we will multiply it by 4, hence 10 bit.

Another way of saying it is there are 256 possible values for each pixel, adding the 4 pixels together make it 1024 possible values, hence 10 bit.

[thesoundone]

Hey, I'm new on the forum, so hi..

 

I just used Thomas' app to convert one of my GH4 files into a .DPX sequence which I brough into davinci, graded and output as 1080p.

 

Its looked absolutely no different to grading the same clip in 5DtoRGB (love this app) and down sampling to 1080p and applying the same grade.

 

Am I missing something? I would genuinely like to know if I am..

 

Cheers!

 

[thesoundone]

Actually, to add to my previous comment, the file that has been through the script is quite a lot more noisy from the tests I've done. Again, it doesn't seem to grade any differently than the original clip. Just my findings.