Red compression and matrix tests

[Muttondraw]

Geoff has posted the final set of images over on Cinematography.net... so that is everything posted.

http://www.cinematography.net/Red/comp-matrix.html

Martin

[WarrenS]

Too bad the links aren't working.

[Muttondraw]

Other images 1, 2 and 3 seem to be linked at the moment, the others aren't yet.

Martin

[Brook Willard]

Now I'm excited. It's like xmas!

[Justin_Kirch]

Starting to key now. I want that REDCODE image to be linked so I can see how I can key that. Also I want to mess with the color matrix grabs.

[Graeme_Nattress]

BTW, I've radically improved REDCODE since those were done :-)

Graeme

[Stephen Williams]

BTW, I've radically improved REDCODE since those were done :-)

Graeme

Graeme,

Any chance of giving Geoff Boyle new pictures? I hope the original materal still exists!

Stephenn

[WarrenS]

Graeme,
Radically is a lot. Tell us more!

[Graeme_Nattress]

The material exists, but as REDCODE is in rapid development, I don't see the point of keep posting up every little change I make. I'll be making more changes soon, so let's wait a bit.

And once the architecture is fixed, there's a large bunch of "magic numbers" that can be altered too....

Graeme

[ericyoung]

The material exists, but as REDCODE is in rapid development, I don't see the point of keep posting up every little change I make. I'll be making more changes soon, so let's wait a bit.

And once the architecture is fixed, there's a large bunch of "magic numbers" that can be altered too....

Graeme

Graeme is such a tease...:laugh:

[Graeme_Nattress]

Sure, but that's half the fun isn't it? Don't worry, it's not just you, it's Jim and the rest of the team that I tease too!

Graeme

[Justin_Kirch]

Was keying in AE in the 4k film comp settings and I just had to point and click, nothing was a problem. I saw all the detail in the mesh trash can and no color bleeding of the greenscreen. This is awesome!

[Emanuel]

Sure, but that's half the fun isn't it? Don't worry, it's not just you, it's Jim and the rest of the team that I tease too!

Graeme:laugh:

[Y.Goh]

BTW, I've radically improved REDCODE since those were done...On the side note and related to the RED's camera: If the matrix is change/modify to allow for more dynamic range, then would the gamma curve be more compressed as a trade-off?

[Graeme_Nattress]

We're not changing matrix or gamma curve to get more DR - we're doing some voodoo sensor magic. I think "matrix" was just a turn of speach.

Graeme

[WarrenS]

Graeme,
Jim said you made a significant "discovery" today. Care to enlighten us??

[Y.Goh]

We're not changing matrix or gamma curve to get more DR - we're doing some voodoo sensor magic. I think "matrix" was just a turn of speach.Speach?? Is there a fundamental difference between modifying the matrix and doing some 'voodoo sensor magic'? :)

[Brook Willard]

This is fun. It's wretched and took about 4 minutes, but the key is stunning at 4K. One click, one slider and it's clean as hell. A little tweaking and color work [and a background that isn't old, poorly lit and matches] and some amazing stuff could be possible.

I'm so excited. I just threw a second green screen image into the same tree and got the second [coffee] image.

[inthemind]

WOW! Brook, that looks great.
I'm so excited to play with this new camera, but I doubt I can light things as nicely as a Pro gaffer.

[Y.Goh]

We're not changing matrix or gamma curve to get more DR - we're doing some voodoo sensor magic. I think "matrix" was just a turn of speach.Thanks for the e-mail help, you've been responsive and knowledgable. Whatever you or the RED's team do to the sensor, DR to me is very important, as I'm doing lots of scene that I have no control of the lighting.

[Chris Kenny]

Keyed with Keylight in Shake, with no tweaking at all (I just set the screen color). 100% crop. The hair is pretty amazing.

Heh. I've been doing some keying on SD footage (from a 16mm original). It hasn't been fun. I think I'm going to like green screen work with Red a lot more!

http://www.nicedissolve.com/images/key.jpg

[Justin_Kirch]

Very cool Chris. I'll have to post my grab up pretty soon.

[firewallet]

Very nice Chris, looks great!

[Jannard]

Chris... nice work. Did you try the basket shot?

Jim

[Brook Willard]

http://homepage.mac.com/brookwillard/basket.jpg

Another one-click key.

[Chris Kenny]

Chris... nice work. Did you try the basket shot?

Yeah, very impressive where the mesh of the basket interacts with the hair, and the key still works with no tweaking.

If every camera was as good as yours, and every DP could light a green screen as well as David Stump, life would sure be a lot easier in post!

If Redcode-compressed footage keys as well as this uncompressed stuff, I'm going to start seriously considering some more effects-heavy projects I wouldn't have previously considered attempting. Though the render times for 4K keying scare me a little.

[Emanuel]

I'm without words Brook! :dankk2:

EDIT


Chris... nice work. Did you try the basket shot?

JimJim,

Simply amazing, my friend! Sorry I'm daring to call you so, besides and beyond our contacts over these forums BUT you can just be, after what I've been finding coming from these same boards...

E.

[firewallet]

I'm without words! :dankk2: Brook!

E.

Same here Emanuel :thumbup: . thanks for the "one click" demo Brook for what should be a nightmare key. :)

RED Team you all keep raising the bar with your vision, hard work and execution. :dankk2:

BTW for an ex-military guy that has traveled a bit... nicest basket trick that I have ever seen... :kiss: (yeah, I said it!) haha

[boothba]

Too late - everybody has already had a go. Oh well here's one more...

[Emanuel]

BTW for an ex-military guy that has traveled a bit... nicest basket trick that I have ever seen... :kiss: (yeah, I said it!) hahaLOL :)


Too late - everybody has already had a go. Oh well here's one more...Don't say it! :) You have yours, it's your right -- maybe having the basket magnified? It would be nice to compare... :beer:

[boothba]

Here's some CUs.

[Brook Willard]

Now that's nice! What app? The softness really sells it...

[Emanuel]

:dankk2: boothba!

[boothba]

Now that's nice! What app?

Inferno - lucky me... except I'm supposed to be working... Don't tell the boss. :(

P.S. $2 to whoever can guess where I stole the bg. hint: it starts with dprevie......

[Brook Willard]

DPrevieQ???

[boothba]

Judges say no. Close but no $2.

[Brook Willard]

You inspired me.

http://homepage.mac.com/brookwillard/compsthumbs.jpg
Clicky (http://homepage.mac.com/brookwillard/comps.jpg)

They're by no means final... but damn. It's just easy now.

[Emanuel]

I was @wrong_thread posting this. Sorry folks! Now it's right on the way! :)

Here are two grading downrez and cropped rough samples taking advantage from boothba post, David Stump lighting setup and, of course, RED ONE capture recording for greenscreen purpose:

more magical hour setup
http://img176.imageshack.us/img176/7842/bucketheadhdwidecropcc4aq9.jpg (http://imageshack.us)

more pre-dusk time (maybe later than the previous one -- who knows? :))
http://img166.imageshack.us/img166/3962/bucketheadhdwidecropcc5qj9.jpg (http://imageshack.us)

[Emanuel]

http://img180.imageshack.us/img180/1707/bucketheadhdwidecropcc5vh9.jpg (http://imageshack.us)

[boothba]

Very nice Emanuel. You found my comp's balls. :dankk2:
I suspected they were hidding there somewhere.

BTW - ya know, I must say those are starting to look pretty damn filmic.

[boothba]

You inspired me.

http://homepage.mac.com/brookwillard/compsthumbs.jpg
Clicky (http://homepage.mac.com/brookwillard/comps.jpg)
.

Ha ha... I'd recognize those backgrounds anywhere. I smell some legal action brewin' at dprevieQ.

[Brook Willard]

Heheh, I claim no credit. Hell, I didn't shoot a single thing in that post. It's all too crispy but it still works.

[Emanuel]

I was @wrong_thread posting this. Sorry folks! Now it's right on the way! :)

Here are two grading downrez and cropped rough samples taking advantage from boothba post, David Stump lighting setup and, of course, RED ONE capture recording for greenscreen purpose:

more magical hour setup
http://img176.imageshack.us/img176/7842/bucketheadhdwidecropcc4aq9.jpg (http://imageshack.us)

more pre-dusk time (maybe later than the previous one -- who knows? :))
http://img166.imageshack.us/img166/3962/bucketheadhdwidecropcc5qj9.jpg (http://imageshack.us)BTW your latter pics Brook, I'm watching by now (and at the same time) a Mel Ferrer late '50s movie [LINK (http://www.imdb.com/title/tt0052864/)] on TV. And I can assure after seeing a greenscreen (Anthony Perkins on the boat going up the river) if 35mm or not, hollywood or not, year '59 or not, their option (aesthetic option not exactly techie one) for a deep DOF is less cinematic what an indie can get with the shallow DOF from the RED S35 sensor. :) (&) Such outcome wouldn't/won't be only coming from the studios!

[im.thatoneguy]

Now that's nice! What app? The softness really sells it...

I was actually going to say it was a little too soft but other than that was very good. Maybe 1 pixel too soft.

Either way 'Claps and Giggles' all around. I'll have to take a whack at it tommarow when I have some time off.

[Brook Willard]

I was glancing through the thread again and noticed a few [many] things wrong with my composites. One of the most irritating parts was this tiny little one-pixel outline that was visible in certain areas but not others. I pulled up the script and zoomed in on it to see what the hell was going on.

Oh. It was cleanly keyed arm hair.

[edit: JPG compression killed this one... it wasn't just blurry "fuzz" in the 16-bit comp.]

[scienceguy_ae]

I decided to see how well Motion's Primatte could pull a key on this one. Not bad...

[Shaw]

Damn, I want a RED camera :(

Just imagine the sort of visual effects low budget guys are going to be pulling off with this stuff and Shake.... I'm envious.

[boothba]

I decided to see how well Motion's Primatte could pull a key on this one. Not bad...

Very nice. I gotta get my hands on some Primatte action. Did you have to mess with that key much? It handled the hair and the guys arm very well.

[PixelMagic]

Can someone create mirrors? I am a student in college studying to be a feature film compositor, and I'd like to try some of these composites myself.

[Brook Willard]

The only 3 images that are available are available at good speeds from the link in the first post.

[scienceguy_ae]

Very nice. I gotta get my hands on some Primatte action. Did you have to mess with that key much? It handled the hair and the guys arm very well.

I did spend just a minute or two fine-tuning it, but not a whole lot. It was surprisingly easy.

I can't wait to get my hands on some footage.

[Brook Willard]

That makes two of us. I really wish they'd post some 2K footage... something beyond 8-bit if possible. I know it'll surface in due time. :)

[Emanuel]

I was glancing through the thread again and noticed a few [many] things wrong with my composites. One of the most irritating parts was this tiny little one-pixel outline that was visible in certain areas but not others. I pulled up the script and zoomed in on it to see what the hell was going on.

Oh. It was cleanly keyed arm hair.

[edit: JPG compression killed this one... it wasn't just blurry "fuzz" in the 16-bit comp.]Following your hint Brook and taking advantage of the scienceguy_ae pic, here are my finds:


http://img237.imageshack.us/img237/2000/rgbscene27comp1jpgcrop1ue2.th.jpg (http://img237.imageshack.us/my.php?image=rgbscene27comp1jpgcrop1ue2.jpg)


http://img153.imageshack.us/img153/3980/rgbscene27comp1jpgcrop2pp1.th.jpg (http://img153.imageshack.us/my.php?image=rgbscene27comp1jpgcrop2pp1.jpg)


http://img237.imageshack.us/img237/518/rgbscene27comp1crop3uprhg3.th.jpg (http://img237.imageshack.us/my.php?image=rgbscene27comp1crop3uprhg3.jpg)

[scienceguy_ae]

Following your hint Brook and taking advantage of the scienceguy_ae pic, here are my finds:

Yes, I also noticed that Motion didn't handle the arm hair as well as Shake. Still, it's not bad for what can be a close to real-time key.

Another drawback to Motion's current GPU based rendering is the 2,048 x 2,048 limit. I had to scale the entire frame to 1920 x 1080 before compositing. Shake handles much larger content.

[edit] I just realized one reason why the arm hair didn't key out as well in Motion. I was working at 1920x1080, as opposed to Brook's 4k comp. Duh.

[Brook Willard]

Yeah, I was 4K in shake. I used Keylight instead of Primatte - I find it to be more effective with "perfect" key images. It's also a very effective one-drag keyer.

[Y.Goh]

Brook, is that you in the picture holding the trash?

[Brook Willard]

:laugh: :laugh: :laugh: :laugh: Don't I wish! I've had a few people ask me the question [while I was working on the key], but no... that's somebody from the RED team or an actor.

Though - to Jim et. al - I will happily donate my services if you need somebody for future shoots. It's a 10 minute drive :laugh: :thumbup:

[Mr. Blonde]

The basket shot is especially impressive! God I'm so damn excited to get my hands on a Red....

[jaysedai]

Good 'ol AE 7 + Keylight (1 click) + Levels. First photo downscaled to 25%, second close up downscaled to 75%.

Jeremy - Red #63

[PixelMagic]

Ok, so I ran a few compositing tests at 2k. Here are the results.

http://img179.imageshack.us/img179/1544/redcomp01xn2.th.jpg (http://img179.imageshack.us/my.php?image=redcomp01xn2.jpg)

http://img182.imageshack.us/img182/1205/redcomp02hk8.th.jpg (http://img182.imageshack.us/my.php?image=redcomp02hk8.jpg)

http://img151.imageshack.us/img151/2398/redcomp03ce7.th.jpg (http://img151.imageshack.us/my.php?image=redcomp03ce7.jpg)

[Haakon]

Wow, you guys are getting way better keys than I am... or does Photoshop just suck at this task?

[jbeale]

Photoshop is nowhere near as good as the real keying tools, at least when I'm trying to use it!

> I really wish they'd post some 2K footage...

If Redcode is only 27.5 MB/sec you'd think a few seconds would be feasible, seeing how many 72 MB TIFF files we're downloading. OTOH I haven't heard that the finished codec would ever be made public, let alone the current unfinished version. In the "already existing" field, MP4 is fairly efficient but there would always be questions about what is a compression artifact and what's "real".

[scienceguy_ae]

Wow, you guys are getting way better keys than I am... or does Photoshop just suck at this task?

Photoshop just sucks at keying. Trust me, I've tried. :-)

[boothba]

> I really wish they'd post some 2K footage...

If Redcode is only 27.5 MB/sec you'd think a few seconds would be feasible, seeing how many 72 MB TIFF files we're downloading. OTOH I haven't heard that the finished codec would ever be made public, let alone the current unfinished version. In the "already existing" field, MP4 is fairly efficient but there would always be questions about what is a compression artifact and what's "real".

How about a compressed jpeg image sequence? Or jpeg 2000? Just a few seconds of 2K would be great!

[Sarmoti]

Wow, you guys are getting way better keys than I am... or does Photoshop just suck at this task?

If you just use a magic wand or color range selection, yeah it sucks. There are proper techniques to get better keys, such as using the channel mixer to create an alpha matte or using plug ins from Primatte, Ultimatte or others.

[im.thatoneguy]

I'm now several thousand miles away from my workstation so I can't do any tests... but I still want to :). Would someone do a difference matte on the two images? I can't believe we've gone 7 pages and nobody has performed THE test for compression quality.

Boothba, I'm sure you at least know what I'm talking about.

Thanks.

[boothba]

Boothba, I'm sure you at least know what I'm talking about.

Thanks.[/QUOTE]

Ya - busy night. All our systems are getting upgraded in the morning so I'm in archive hell. I'll try to fit it in.:)

edit - from what I've seen on this thread (and all things considered...) Redcode looks fantastic. Christ, the years I've spent keying blocky HDcam compression..........

[scienceguy_ae]

I'm now several thousand miles away from my workstation so I can't do any tests... but I still want to :). Would someone do a difference matte on the two images? I can't believe we've gone 7 pages and nobody has performed THE test for compression quality.

Boothba, I'm sure you at least know what I'm talking about.

Thanks.

Here's a difference matte showing the compression. I think I did it correctly anyway... I pushed the levels to the extreme to get what you see here, so I suppose the difference really isn't all that much.

Sorry for the small size. Noise isn't easy to compress, but I guess that's the point right?

Does anyone have any idea what those large blocks are?

[scienceguy_ae]

After looking at the compressed image more closely, it looks like REDCODE is removing mostly noise. There are a few spots however, where fine detail like hair is being removed.

Notice how in these two 100% crops of just the Red channel (the noisiest), some of the hair is disappearing with the noise. My guess is that Graeme has tweaked REDCODE even more to reduce this loss. I'm not disappointed at all, just pointing out how good the worst of the compression looks. :-)

[Graeme_Nattress]

Yes, I'm working on making it not eat into the details so much. It's working well so far. And yes, the "artifacts" of wavelet compression are, at light levels, mostly noticed as mild noise reduction. Indeed, wavelet transforms are often used in some noise reduction algorithms.

Graeme

[boothba]

Well I've had a chance to do some difference key tests between uncompressed and REDCODE-RAW and again I am extremely impressed.

As a frame of reference I was hoping to compare the results to HDCAM SR compression but unfortunately I could only get my hands on a dBeta deck. Both formats (HDCAM SR & dBeta) purport to be visually lossless at about 2:1 compression (opinions vary).

For this test I used three images sizes within a 720 X 486 10bit frame. The first image is a 100% crop of the 4.9K image, the second image is a 50% scaled crop (displaying 2K resolution) and the third image is about a 18% scaled crop (RED @ SD). I then difference keyed the uncompressed with the REDCODE image and compared that with a dif of the same image output to dBeta and re-input.

The first thing I noticed is that our dBeta deck needs a cleaning! :) The next thing to note is that REDCODE becomes more invisible when you scale down, to the point where it is (pixel for pixel) significantly better than dBeta (@ 2:1). This is great news for the tiny minority of us who will NOT be screening our films on a 4K projector - at least not for a while. But 4K aside, REDCODE should be able to deliver the finest 2K & HD commercially available to us plebs who can't afford a RAID. I will attempt to repeat this test when I get my hands on a HDCAM SR deck for a more official comparison.

Two things to remember:

ONE - I used the same difference key setting for all images, but the parameters are seriously cranked - the human eye would never see these anomalies.
TWO - dBeta is an widely accepted industry standard format. In North America the vast majority of SD shows are mastered onto dBeta (no BetaSP jokes). In fact a great many post shops still use dBeta as an archive format and broadcast content is often dubbed numerous times before hitting the air. Good to remember when you look at the bottom right image.
___________
Alex Boothby

[scienceguy_ae]

Yes, I'm working on making it not eat into the details so much. It's working well so far. And yes, the "artifacts" of wavelet compression are, at light levels, mostly noticed as mild noise reduction. Indeed, wavelet transforms are often used in some noise reduction algorithms.

Graeme

Hmm... noise reduction and compression, all in one pass. Nice. :)

[Graeme_Nattress]

With DigiBeta, what you're mostly seeing are the problems ofRGB to Y'CbCr 4:2:2, which does some quite nasty things to edges. And what you're not seeing is the 12bit to 10bit, but that's trickier to see, but once you get the data into REDCINE, it makes the world of difference.

Graeme

[jack78]

I think the images from the Mysterium sensor look very good. One thing I’ve noticed though is that the debayer algorithm is doing somewhat odd things with hair.

http://img299.imageshack.us/img299/5418/hair100cropxu5.png

[scienceguy_ae]

I think the images from the Mysterium sensor look very good. One thing I’ve noticed though is that the debayer algorithm is doing somewhat odd things with hair.


Really? It just looks a little soft to me, probably due to the lens not being perfectly perpendicular to the sensor. What odd things are you referring to?

[jack78]

I can’t quite put my finger on what it is. The hair looks rough instead of silky smooth.

[scienceguy_ae]

I can’t quite put my finger on what it is. The hair looks rough instead of silky smooth.

Maybe she just needs to switch conditioner. :D

[Graeme_Nattress]

Take two HD cameras into the shower? Me, I just shoot 4k on RED and go!

Graeme

[jack78]

:laugh: All right, I’m probably seeing ghosts.

[Sarmoti]

Difference maps show the CODEC compromises mostly noise and very small detail. Looks flattering on talent too. REDCODE images are going to scale and compress very nicely for HD delivery. If digitally projected, I think that I'd be inclined to add digital grain, not a bad problem to have... Congrats, very impressive so far.

[Sarmoti]

Well, there's yer problem...

http://www.matthewgreene.com/misc/hairzoom.jpg

P.S. Had to use the uncompressed image... REDCODE got too blurry to tell if it's male or female...

[Brook Willard]

Now THAT'S great.

[Christopher Barry]

Sarmoti, that is awesome. :D Can RED do the same as in Blade Runner, see details around the corner, behind a door frame?

[Beat Takeshi]

LOL!
I thought i saw something in there.

[im.thatoneguy]

Thanks to all who obliged my request.

I think the large blocks in the first image are optimization patches. Graeme is probably coding it to isolate areas which have finer detail, in order to compress them differently from large even areas. Almost like an OctTree compression scheme. Which is really a good hybrid between your normal block encoders, except he's encoding within a basic block compression scheme.

My opinion on Jack's image isn't a focus issue. If that's from the uncompressed it's probably just a bunch number of hairs falling in an awkward place for the de-mosaic algorithm as was originally suggested.

Graeme I would pose one question to you: when I compress a JPEG2000 in photoshop there are two compression schemes, one of them generates funny little + marks. One of them always creates a clean image. Sorry I can't remember which is which (and again no workstation nearby). But I think the "Fast" one doesn't generate the + marks and actually creates in my opinion more accurate highly compressed images... are those visual + marks intersections of optimzation boxes? I can see one of the + in one of the first images, where some crankes the contrast on the compressed image... it worries me more than softness just because it's actually a visual artifact that looks digital. Any idea what those are? How to optimize those? Edit: (http://www.nicedissolve.com/images/redcode2.png) You can see them here.

[jack78]

Well, there's yer problem...

http://www.matthewgreene.com/misc/hairzoom.jpg
See? I told you guys there was something with the hair :laugh: