Song Of The South - many projects, much info & discussion thread (Released) — Page 23

digitalfreaknyc said:

You really should do 1080p. 720 is almost archaic in this day and age.

Agreed.

Doctor M said:

I can agree with that logic.

Most sources say a 4k scan is 100% of the detail of a full size print.  I'd think a 16mm has half that.

A 16mm or 35mm negative can have more than 4K of data, but a print never really does.

A good 16mm print is lucky to have 1080P worth of *real* data, and the dynamic range tends to be crushed as well. A reduction print is lucky to offer anything much above SD resolutions and dynamic range.

However, you want to work at higher than your delivery resolution whenever possible.

i.e. if you want to deliver a 720P 8bit print, then work in at least 1080P 10bit when doing your edits/correction etc.

If you want to deliver a 1080P print, then ideally you would do all of the editing work in 4K. Want to deliver in SD, then work in 720P.

Everything you do to it causes some loss, stabilising, colour correcting etc. so you want to work in the highest resolution/bit depth you can to avoid data loss and have the ability to work at a sub-pixel level.

Then for the final output you render at the lower resolution.

With your print, 1080P will be more than it holds, so I'd work in that resolution.

Remember it is not a scope print, (from memory it is 1.37:1) so the actual resolution you will be working with will be more like 1450 x 1080 or thereabouts.

I'd be aiming for a 720P release, but work in 1080P (i.e. 1080 lines).

You can always output then in 1080P if you want, but I don't think the data is there.

Just make sure you have lots of HDD space, at 1080P, 16bit per channel colour, 1 minute of footage weighs in at about 3 gigabytes, you can do the math from there :)

poita said:

Just make sure you have lots of HDD space, at 1080P, 16bit per channel colour, 1 minute of footage weighs in at about 3 gigabytes, you can do the math from there :)

Is this before or after lossless-compression?

That is compressed.

1920x1080 = 2073600 pixels per frame

16 bit colour per pixel for RGB = 48bits per pixel total.

2073600 pixels x 48 bits = 99532800 bits

99532800 bits /8 = 12441600 bytes

=12150 kilobytes

= approx 12MB per frame uncompressed.

At 24fps for film, 1hr of footage = 60x60 seconds = 3600seconds

3600 seconds x 24 frames per second =86400 frames per hour.

86400 x 12MB = 1036800

= 1037GB per hour uncompressed.

So around two thousand gigabytes for a 2hr movie.

So with lagarith you end up with around 300GB per hour or so, with MSU about 200GB per hour, depending on how much grain there is etc. So around 300-500GB for a typical feature film.

Halve those numbers for 8bit RGB (24bit colour), cartoons compress better than action films etc. etc.

I am running on no sleep at all, so any or all of the above math could be wrong, but the lagarith compressed sample clip is around 100MB for less than 3 seconds, so it would come out at around 150GB per hour, and I think it is 8bit from memory, so it sounds about right.

Images from Reel1:

https://www.dropbox.com/sh/7taojpt745e7t6l/fUe-ItbBgF

poita said:

Images from Reel1:

https://www.dropbox.com/sh/7taojpt745e7t6l/fUe-ItbBgF

Nice. Thanks for sharing!

We'll probably have to do three passes to correct the colors:

1. Custom curves for each reel. That is, a single setting to separately bring each entire reel to an acceptable state; e.g., reel01.GiCoCu(...)

2. A shot-by-shot color-match against ww12345's restored BBC broadcast using the N-Dimensional PDF Transfer MATLAB function. Question: How fast does this work on your machine?

3. Shot-by-shot manual tweaks by an experienced colorist with some good references.

#1 seems simple enough and despite being long and tedious, #2 seems simple enough as well. Who can we find to tackle #3?

ww12345 said:

As far as number two, is that a question directed at me? If so, what do you mean? :) 

I'm sorry! It was meant for poita.

The matlab code isn't particularly speedy, but there is nothing specific to matlab as far as the math goes. It should be possible to recode it in any language and get much much faster results.You would want to do this if intending to deal with that amount of data.

Matlab is great for theorising until something works, and then get it re-written in something else a little more swift once you have the concept proven.

jimjimmyjones85 said:

I am very glad to see the progress you are making. I am excited about your progress and the movement you are having.

and I thought only my doctor was interested :D

Why would you specifically use ww12345's to match colors to?

I believe he modified them during the conversion to NTSC.
Mentor's untouched release is probably a more accurate source.

Btw, saw the screenshots and hope you find some color in there besides red.

Actually that's only the left eye of a long lost 3D version.