Video pre-processing - Page 4

Nope the process I've described entail sub pixel filtering in both x and y. I don't see any ringing on this transfer it's just mightily soft in comparison to the 2k.

The ratio doesn't stay the same with a letterboxed 2.35:1 film. When you crop on all 4 sides you're basically zooming in therefor the active picture area is larger within the 16x9 frame.

Its fine Kram for all the reasons I've stated . It represents a smaller error tolerance than many of the previous stages including the initial capture.

This is 2048x1556 ( or 2048x1152 if its 3 perf) not 2048x1080.

Doing some further work tonight its definitely a crop to 1920 from 2048. The practical difference of this is that most of the cropped area is soft because its at the periphery of the frame and mostly vignetted. Its also well within the tolerance defined by the lineup slate for the feature. I'd take a bet that there was no additional horizontal info on display in the theatrical release vs the BD.

Further to my previous comments I can confirm the crop and resize is exactly the same comparing shots from scenes very far apart in the movie and originated at 3 and 4 perf : 1920 crop and a 0.984 rescale down.

I'd probably revise my description of the transfer: it is softer than the 2k , noticably so in direct comparison to the 2k but I don't see any increase in aliasing or sharpening artifacts. So in those terms its a very clean transfer its just not up to the sharpness of the 2k which I'm surprised and disappointed by

There are obvious artifacts from the component downsampling but the rescale I'm seeing itself may just be down to the filtering on the component downsample or upsampling (in the playback software) stages.

I can do my own 4:2:0 component conversion and back to RGB without displacing the pixels to this level though. I'm obviously not having to deal with any compression though.

Levels are fine I ran some test patterns through it first and checked the values. I'm comparing 10bit log 2k with jpegs off 8bit 4:2:0 highly compressed video so any variance in color is somewhat moot given the differences already involved.

I am ball park grading the 2k to match the video for easier comparisson. I will not be able to make a direct match without trashing the 2k which kinda defeats the point.

How does luminance alone compare which does not nee to be downfiltered?

You are zooming without resampling. Losing edges that often have issues anyway is no big deal. Films are not shot to have relevant detail at the very edges.

Soft but the grain structure is intact no signs of DNR.

2048x1556 image of alternating black and white pixels ( worst case scenario for filtering).
Downscale is the sinc filter in Shake

http://www.avsforum.com/avs-vb/attac...1&d=1249898407

http://www.avsforum.com/avs-vb/attac...1&d=1249898407

I wonder if the magic number 0.984 is something to due with the "clean aperture" of 1888x1062? 1888/1920 = 0.9833333

John

You know it could be but I would have thought any edge transition filtering malarkey would have been catered for already by doing a straight 1920 crop out of the 2048.

Unless its something they've borked at the encode.

That's an interesting website you've got in your sig John.

I agree the resize look crazy, my guess is somebody trying too hard to keep to "the standards" and not letting some tolerance in one area avoid borking some other. I doubt it's an encode thing but it could be, how does the encode process work in practice, who sends what to whom and in what format? My mental picture has an uncompressed 4:2:0 1080p 24ishHz master going off to the next step, is that even close to what really happens.


Sadly it's not very interesting yet
Once I've got something half decent working then maybe.

John

I'm not entirely sure where the downsample to 4:2:0 happens. I suspect its handled at the encode from at least a 4:2:2 source. D1 4:2:2 used to be commonly handed to an encoder for dvd generation. It would make sense that they encode from a larger pot into a smaller one unless there are any over-riding advantages to handle the 4:2:0 downsample prior to encoding.

Maybe a compressionist could advise?




Seems like a very interesting and worthwhile project though , look forward to it.

It would also be interesting to know where in the chain the 3:2 pulldown info is added



Thanks

That should be in the encoder, since the actual compressed file is 24P with flags.

Thanks stacey

Looks like the real world is a bit messy, shouldn't have expected anything less I suppose.

I'm still a bit unclear how the bits get from the grader to the encoder.

I think the process from the encoder to the disc is fairly clear. Getting the image off the film and into digital form seems to be reasonably standard but at the moment based on the above discussion I'm not clear what it is that the grader passes to the next step and how many steps there are before the encoder. It seems that ideally the grader would output uncompressed 4:2:0 1080p, but I suspect the real world is often messier than this ideal.

John

Someone feel free to intercede if I'm wrong.

To my knowledge, film is typically graded in 2K DPX 10-bit log. The finished look is then (typically) output to tape like HDCAM-SR 1080p 10-bit 4:4:4 or a SAN. From there it can be ingested onto the local encoding machine via capture card as 10-bit 4:2:2 and encoded.

couple of points

The DI is usually 10bit log (DPX fles don't have to be log..or 10bit though) but the film also has to be graded into video colorspace too ( analagous to a telecine in the old days) sympathetic with the film look but compromised accordingly into video.

The video master is not the same as the DI. I've read some comments recently from people thinking they get the DI on BD albeit compressed at 8bit. Apples and Oranges.

They don't technically "capture" the material off HDcam , they just grab the data off the tape ( same as D1 in the early days of dvd.....yuv files I think)

1:1 crop from the opening seconds of the two Blu-rays of Watchmen


Is the chroma shift an inherent issue with 4:2:0, or is it actually misaligned in the left image?

I'm also wondering about the resolution loss and ringing on the right image, as mentioned earlier. Not sure whether that's OT for this discussion though.

Which versions are each from by the way?

chroma1: luminance is clean , color difference channels definitely have filtering issues . I'd say its down to the 4:2:0 downsample being problematic , it may be more visible on this material because of the strong yellow. Its more like a smeary filtering effect than a misalignment.

chroma2 : all channels seem clean but there is ringing on all three so I'd say they added this prior to the downsample to 4:2:0 as the ringing is suitably mushy in the u and v channels.

You get nice results if you combine the y from the first image with the u and v from the second: maybe we should go analogue component and swap the channels around as required from multiple versions

Left image is USA Blu-Ray (Director's Cut). Right image is UK Blu-Ray (Theatrical Cut). The right image is IMHO resampled in X direction, but not in Y direction. Check out this comparison shot:

http://comparescreenshots.slicx.com/...1125/picture:2

It's clear to see (check out the detail in the windows in the middle of the frame) that the USA Blu-Ray is horizontally razor sharp, while the UK Blu-Ray is really soft in comparison. I think we can see here what Mr.D has tried to make clear, namely that cropping the 2K master produces much sharper results than rescaling the image. Unfortunately the USA Blu-Ray has a much lower bitrate and was DNRed. So neither Blu-Ray is really good.

Hmm curious not entirely sure what I'm seeing here.

Apart from the opening logo above the UK version doesn't appear to have any obvious ringing or DNR and has good grain structure. I have to say I prefer it over the mushy US version. I don't really see the horizontal filtering on every shot ( I see it on the windows) and I would have expected it to impact the grain structure to a degree. I'm wondering if its an encoding artifact rather than a resize issue.

If one was a crop and one was a repo I wouldn't expect the transfers to line up pixel for pixel which they seem to so I'm assuming the differences we see between these two are down to the compressionist.

I much prefer the UK transfer overall. I'm assuming no resizing or repoing was done to make these two versions line up with each other in the examples posted?

The issue I was posting about is slightly different as I'm essentially comparing the BD against the original 2k scan.What I'm seeing is that the BD transfer is a 1080p crop out of the 2k not a resize but it also has this tiny weird little non-centred resize.

Agreed.


Possible, I don't know.


Don't know that, either, the comparison was not done by me.

Correct, no manipulation was done to the output from ffdshow aside from converting to RGB and adding the overlay text identifying the versions.

Some more pics I posted in my Watchmen thread.


By the way, this reminded me of something that pissed me off a few years ago...




Both are directly copied DVB-S streams, one is from 2006-11-12, and the other 2006-11-18. Sometime in between those dates, the channel started looking like image #2 all the time...

Used to be the best-looking SD channel available around here. Any idea what would cause something like that?

Short answer: corporate greed. :-/

How is dither pre-processing handled when you have a grainy source? Is it even needed? Does the grain (a sort of dither if you will) affect how the algorithm operates?

Mr. D:

Could you evaluate/comment on this article please:

"HD" versus "SD" color space

http://www.glennchan.info/articles/t...olor-space.htm

What part exactly Lee? Looks reasonable from scanning through it.

Just the overall accuracy - thanks for taking a look at it.

It is interesting that the areas of fine white lines in the windows and the areas around the bright lights on the buildings change their overall brightness between the 2 versions, looks to me like a symptom of some processing being done in gamma adjusted space but I'd probably finger the noise processing on the US version.

John