F3: 444 vs. 422 Anyone had experience comparing the two?

[starcentral]

Ok just to clarify, my take is that for true 444 color space you need a dedicated photosite for each R G B value for each pixel. There can be not much debate here in my opinion.

But for resolution it makes sense that the amount of sampling is not limited to the sensor pattern itself, ie. this is why 422 is "better" for keying over 420. Now strictly speaking 444 sampling, it's possible on any sensor of course and might yield some increase in resolution but isn't going past 422 the point of diminishing returns?

Lastly I wonder how the differences might compare between RGB stripe sensor, Bayer Pattern sensor, and the new Q67 diagonal oriented photosite sensor as found in The F65?

 

[Razz16mm]

So I'm assuming that once you start going past 4:2:2 10-bit, the returns are not as great as jumping from 4:2:0 8-Bit?

So is the jump from 4:2:0 8-bit to 4:2:2 10-bit really that big?

The jump from 4:2:0 8-bit to even 4:2:2 8-bit is huge in terms of grading range. Especially when the 4:2:0 is a product of a highly compressed long GOP codec like AVC/H.264.
Playing around with grading S-log AVC internal recording sample from an FS700 it barely had enough data to bring it to normal contrast with a slightly desaturated look. Pushing saturation much at all resulted in bad macro blocking on high contrast borders and banding in low contrast areas. On the other hand grading highest quality 8-bit 4:2:2 240 mbps broadcast AVI files produces no such artifacts and fine differentiation in tonal values. The higher the data rate in an I-frame codec the more room you have to play with whether 8-it or 10 bit. Compression is the enemy of flexibility.

 

[starcentral]

I think most people here are saying that 422 over 420 give more bang for buck in the line edge department, and not color. To get more color, or graduations of each tonal range for R G B you'd really need to go from 8-bit to 10-bit so you can go from 256 levels to 1024 levels (4 times more) in post.

Any thoughts on that?

 

[Razz16mm]

I think most people here are saying that 422 over 420 give more bang for buck in the line edge department, and not color. To get more color, or graduations of each tonal range for R G B you'd really need to go from 8-bit to 10-bit so you can go from 256 levels to 1024 levels (4 times more) in post.

Any thoughts on that?

Compression below the broadcast recommended minimums of 50 mbps and chroma sampling reductions seem to have far more damaging effects on post grading flexibility than the difference between 8-bit and 10-bit. 10 bit certainly gives more refined control, but won't help if your starting point is sub 30mbps 4:2:0 AVC. There is no data there to work with even if you transcode. Transcoding Sony Slog-2 AVC to 240mbps 10-bit 4:2:2 AVI broadcast transport codec didn't help the grading at all.

 

[KINOKS]

What matters is: 1. compression, 2. bitdepth, 3. noise and 4. colour resolution. Try grading a 100Mbs 420 8 bit footage and compare it to 100Mbs 422 8 bit footage. Preferably try grading the same shot. You wont see difference.

...

 

[starcentral]

I pushed these final colors out of some footage shot using a middle of the road white balance (4400k) in 8-bit 420 XDCAM on the F3.

While in motion the quality is not too bad for YouTube where some image degradation can become hidden, but on a still frame you can definitely see the posterization on any gradients such as the forehead, left hand near watch, and shadow on couch.

So would more higher color sampling have really helped better this image over bit-depth? I can see posterization even on a still frame from the original so I see the bit-depth as being limiting not the color sampling.

Before Post

After Post (click to enlarge to original)

 

[Razz16mm]

4:2:0 tends to bleed at high contrast color boundaries and get cross color contamination issues with highly saturated colors. Below is an 8-bit 4:4:4 color sample. Shading of skin tones is much more detailed and refined from the warmer key to the cooler shadow side. But in general I would agree with Sanjin's statement above.

 

[maarek]

Pushing saturation much at all resulted in bad macro blocking on high contrast borders and banding in low contrast areas.

Though to be fair, that has nothing to do technically with 4:2:0. It has everything to do with low bitrate. Just take a jpeg shot, encode it with 100% quality, other at 4:2:2 and other at 4:2:0.

Push around. (the real differences are in high contrast color edges or very saturated color edges), there will ne difference in macroblocking or banding. 4:2:0 is used because it lets the bitrate go a bit lower before showing compression artifacts.

 

[Razz16mm]

Ok just to clarify, my take is that for true 444 color space you need a dedicated photosite for each R G B value for each pixel. There can be not much debate here in my opinion.

But for resolution it makes sense that the amount of sampling is not limited to the sensor pattern itself, ie. this is why 422 is "better" for keying over 420. Now strictly speaking 444 sampling, it's possible on any sensor of course and might yield some increase in resolution but isn't going past 422 the point of diminishing returns?

Lastly I wonder how the differences might compare between RGB stripe sensor, Bayer Pattern sensor, and the new Q67 diagonal oriented photosite sensor as found in The F65?

I've had some rounds with Barry over this, but you can't really compare encoded color spaces directly with the sensor sampling pattern and gamut. Stripe gives you equal samples for all three values at 1/3rd of sensor resolution. Bayer gives you a sample pattern and distribution that mimics the human eye with a 2:1 spatial sample of green to red or blue plus 100% luminance sample at full resolution. Q67 is still Bayer, just rotated 45 degrees. What you can get in ultimate color resolution is determined by the sophistication of the interpolation and encoding processes used and how much the sensor over samples compared to the encoded output resolution. But any way you cut it, less than 4:4:4 encoding discards some of the original color data the sensor delivered.

 

[Barry_Green]

"mimicking the human eye" has nothing to do with color sampling. Either you have the samples, or you don't. RGB Stripe gives you the samples. Bayer doesn't. There's really not much more that needs to be said about it. You cannot, now or ever, get 4:4:4 color sampling out of a Bayer sensor at anything higher than 1/4 stated resolution.

Talking about the human eye is just a complete red herring that has no relationship to digital color sampling. The human eye is profoundly easily fooled. It's analog, and it's subject to however the brain chooses to interpret what it sees, and it results in some truly amazing optical illusions. Is that what you want in your digital video imaging system? It sure isn't what I would want in mine.

Bayer is a mediocre solution to a problem that has no easy solution. It's not an ideal or an end goal. There are tradeoffs in all systems -- Foveon involves tradeoffs, but true 4:4:4 color sampling. Three chips involves tradeoffs (primarily, needing 3x as much silicon to do the job) but true 4:4:4 color sampling. RGB stripe involves tradeoffs (lower ultimate luma resolution, but true 4:4:4 sampling at its intended final output size). Canon's C300 system involves tradeoffs (2K/1080p output at 4:4:4 instead of 3.2K de-bayered 4:2:0-ish).

They all are compromised systems. Pick your compromise. But just acknowledge what the compromise is and don't try to bring "magic" or "pseudo-science" into it. Bayer is a way to simulate a color-sampling sensor by synthesizing information, not by sampling it. The demosaic process takes a 1/4-size blue (or red) sample and up-rezzes it to full size, through synthesizing the missing information. There's no reasonable way anyone could call that 4:4:4 "color sampling", because only 1/4 of the relevant information was sampled, the other 3/4 was synthesized, guessed, predicted... but not sampled.

 

[Shoots56]

"mimicking the human eye" has nothing to do with color sampling. Either you have the samples, or you don't. RGB Stripe gives you the samples. Bayer doesn't. There's really not much more that needs to be said about it. You cannot, now or ever, get 4:4:4 color sampling out of a Bayer sensor at anything higher than 1/4 stated resolution.

Talking about the human eye is just a complete red herring that has no relationship to digital color sampling. The human eye is profoundly easily fooled. It's analog, and it's subject to however the brain chooses to interpret what it sees, and it results in some truly amazing optical illusions. Is that what you want in your digital video imaging system? It sure isn't what I would want in mine.

Bayer is a mediocre solution to a problem that has no easy solution. It's not an ideal or an end goal. There are tradeoffs in all systems -- Foveon involves tradeoffs, but true 4:4:4 color sampling. Three chips involves tradeoffs (primarily, needing 3x as much silicon to do the job) but true 4:4:4 color sampling. RGB stripe involves tradeoffs (lower ultimate luma resolution, but true 4:4:4 sampling at its intended final output size). Canon's C300 system involves tradeoffs (2K/1080p output at 4:4:4 instead of 3.2K de-bayered 4:2:0-ish).

They all are compromised systems. Pick your compromise. But just acknowledge what the compromise is and don't try to bring "magic" or "pseudo-science" into it. Bayer is a way to simulate a color-sampling sensor by synthesizing information, not by sampling it. The demosaic process takes a 1/4-size blue (or red) sample and up-rezzes it to full size, through synthesizing the missing information. There's no reasonable way anyone could call that 4:4:4 "color sampling", because only 1/4 of the relevant information was sampled, the other 3/4 was synthesized, guessed, predicted... but not sampled.

++
Film: the only real 444 option.

 

[starcentral]

^ +++

 

[Razz16mm]

"mimicking the human eye" has nothing to do with color sampling. Either you have the samples, or you don't. RGB Stripe gives you the samples. Bayer doesn't. There's really not much more that needs to be said about it. You cannot, now or ever, get 4:4:4 color sampling out of a Bayer sensor at anything higher than 1/4 stated resolution.

Talking about the human eye is just a complete red herring that has no relationship to digital color sampling. The human eye is profoundly easily fooled. It's analog, and it's subject to however the brain chooses to interpret what it sees, and it results in some truly amazing optical illusions. Is that what you want in your digital video imaging system? It sure isn't what I would want in mine.

Bayer is a mediocre solution to a problem that has no easy solution. It's not an ideal or an end goal. There are tradeoffs in all systems -- Foveon involves tradeoffs, but true 4:4:4 color sampling. Three chips involves tradeoffs (primarily, needing 3x as much silicon to do the job) but true 4:4:4 color sampling. RGB stripe involves tradeoffs (lower ultimate luma resolution, but true 4:4:4 sampling at its intended final output size). Canon's C300 system involves tradeoffs (2K/1080p output at 4:4:4 instead of 3.2K de-bayered 4:2:0-ish).

They all are compromised systems. Pick your compromise. But just acknowledge what the compromise is and don't try to bring "magic" or "pseudo-science" into it. Bayer is a way to simulate a color-sampling sensor by synthesizing information, not by sampling it. The demosaic process takes a 1/4-size blue (or red) sample and up-rezzes it to full size, through synthesizing the missing information. There's no reasonable way anyone could call that 4:4:4 "color sampling", because only 1/4 of the relevant information was sampled, the other 3/4 was synthesized, guessed, predicted... but not sampled.

Nor can you get color sampling at higher than 1/3rd of stated resolution for a striped single sensor.

But again, the Bayer sampling results depends on the sophistication of the debayer interpolation process used which is a 4 vector process: YGRGB, not 3, so while it may not be a full pixel for pixel 3 color sample like a 3 chip camera or a Foveon sensor, it is when processed correctly much better than your assertion it is equivalent to 4:2:0 video because valid blue and red data are recovered from the green pixels based on interpolating the known response slopes of the filters where they overlap. Not magic, just very sophisticated mathematical algorithms. It is not in any way equivalent to encoded video color sampling which discards sensor data the encoder is receiving regardless of what the data is. As far as providing data to manipulate in a post grading process, Bayer raw even at 1:1 resolution is a much richer broader gamut higher color resolution source than any 4:2:0 camera.

Example from a 96 bit floating point raw processor of a 2k Digital Bolex frame, full color and blue channel, not something you will see from any video camera debayer processor as it is too slow for real time motion:
This was a sort of day for night extreme grade look, so the blue frame is darker and less detailed in the shadows than it would be if the scene values were left as in the full color image.