This content has been marked as final. Show 375 replies
Just to get things rolling, I solved my problem posted near the end of Part II - not being able to open clips. Turned out there were two spaces in the file name, not just one. Not sure how that got changed, but all is well, and I an mow testing the newest script.
I have an idea here, Jim:
One of the problems in this type of comparison is objectivity. There is every possibility that (on a given HDTV setup) ALL 24p might look a little choppy (even real film stuff). As we've both mentioned, motion blur and staggered pans are artifacts we see all the time in movies and/or telecined broadcasts, NTSC<->PAL conversions, etc. Some of these setups may be on 60Hz (or even 50Hz overseas) ... others 24/48/96/120Hz, Some are connected with HDMI or DVI, other component, etc. Some DVD players and/or TVs may do a better job than others, upsampling quality differs... blah, blah, blah.
By the same token, it is then possible for a particular conversion setting to give a clip an "illusion of fluity": perhaps more fluid than it was to begin with? So, you wind up with a conversion that looks really spanky on one setup and then looks dreadful on all the others.
There is also the problem that (by virtue of the fact that you've shot this in 60i and you're intimately familiar with the original) you're finding the "best looking" clips to be those that most resemble the original -- not those that most resemble "film".
To remove these variables, I plan to render two versions of a 3D scene of *perfectly* smooth camera and object motion -- one at 24fps and another at 60fps. This way, a real 24p control can be used to evaluate the fluity of motion when the 60fps clip is converted to 24p with my script.
>One of the problems in this type of comparison is objectivity.
Even Jeff admitted the (iM=0, bR=1) setting looked best on his SD CRT, so I think we can get objective, workable results that apply to a wide variety of material and look good on all SD sets. The goal now is to get that same repeatability on HD sets, which do present us with much greater variability. A situation resulting from idiotic technocrats, but there it is.
I think the best way to get there is to have as many people testing on as many different setups as possible. By all means, upload your clips. But I would encourage others to do so as well. Real world footage shot using cinematic techniques. The more, the better.
>you're finding the "best looking" clips to be those that most resemble the original -- not those that most resemble "film".
That, I am quite certain, is not a factor.
>one at 24fps and another at 60fps.
How valid with that be, as real world footage originates at 30 fps, not 60? I would think a 30i sample might be the better test, as it more closely resembles real footage.
>Are y'all using GOPs of 12 N frames between I frames for the 24p MPEG2 files?
I wasn't, but will be now. Good catch.
I don't know, Jim. I can't rule these out completely. I know that we all try to be as objective as possible but it can fail at some point. I certainly try to be objective when I'm writing the script and running my own tests, but I don't always have faith in my own observations either.
> I think the best way to get there is to have as many people testing on as many different setups as possible
It certainly is good, but best when everybody has access to to every setup with every type of material, but that is not realistic. Even if it were, the experiment is hard to control.
EXAMPLE: In the HD -> SD conversion threads, I read post after post of well-meaning people who swore you could get great results by exporting directly as DV, nesting and scaling the sequence, letting Encore do the scaling, etc. I did some controlled experiments and found not only did none of these "look" any better than the other, but they were all different ways of achieving almost exactly the same result.
In our case here, let's take a hypothetical: I run a test converting to 24p. There's a particular segment that looks absolutely perfect on a given setup with particular settings. Perhaps I am so enamored with that one little segment that it causes me to overlook flaws in other segments. The reverse of this is always possible also.
> real world footage originates at 30 fps, not 60
There's no difference between 30i and 60p as far as the framerate conversion is concerned. It really just speeds things up (as there is no need to deinterlace it) and remove the deinterlacing quality as a variable. Apparent smoothness of the frame rate would be the ONLY purpose of the test. Of course, I could weave alternating fields and make it interlaced... but why?
Incidentally, you can use my script with 60p material...
... or another deinterlacer if you wish:
>I don't always have faith in my own observations either.
Agreed, which is why we use community agreement as a guide post to define "objective" with, for now, you me and Jeff being the community, and hopefully with more qualified observers coming on as we go.
>but best when everybody has access to to every setup with every type of material
I was looking more for a few common clips, tested on various setups. If Jeff has some script options he likes for HD, he can post and I can then test on SD. And vice versa. Not ideal, but possibly workable, especially if more are willing to join in and do some rigorous testing.
>and remove the deinterlacing quality as a variable.
I'm not understanding why you would want that in a test, as deinterlacing plays such a vital role in the whole process.
>I could weave alternating fields and make it interlaced... but why?
Because a test of the entire process is more valuable, seeing as that's what folks will end up using. No?
My first short round of tests were done using (Basic=true, DeintMethod=4, bR=1) with the newest script. The clips were converted first, then brought into a 24p project and exported as a 24p MPEG for use in Encore, then watched on an SD CRT.
I definitely saw an improvement in the motion scenes. Pans and tilts seemed smoother, with less ghosting. Overall very film like, more so than my own previous favorite (iM=0, bR=1). However, there were some artifacts not previously present - some flicker on vertical posts, and some sparklies in fine details like tree leaves during panning.
Also, I saw no difference using N=12, but will continue to use it anyway as it just makes sense.
Next, more options...
> I'm not understanding why you would want that in a test, as deinterlacing plays such a vital role in the whole process.
It surely does, but it is best to isolate the deinterlacing and blending as much as possible for testing.
My 3D test would show (given absolutely perfect deinterlacing) which BlendMethod and parameters provide the best appearance at 24p -- and you'd be able to directly compare the "converted" result with the "native 24p" clip, which is not possible with live footage. Got me?
> However, there were some artifacts not previously present - some flicker on vertical posts, and some sparklies in fine details like tree leaves during panning.
I suspect this is the same issue that occurs in DVFilm Maker: the fact that every other frame is unblended will reveal more deinterlacing artifacts AND does this alternating sharp-soft-sharp-soft pattern that can give the appearance of flickering.
To make testing easier, don't even use Basic=true. Use BlendMethod=3 (they are differnt names for the same thing). Do 4 tests like:
dv60i2film(BlendMethod=0, DeintMethod=1, iM=0, bR=1)
dv60i2film(BlendMethod=3, DeintMethod=1, iM=0, bR=1)
dv60i2film(BlendMethod=4, DeintMethod=1, iM=0, bR=1)
dv60i2film(BlendMethod=5, DeintMethod=1, iM=0, bR=1)
Another thing, Jim (and I lied to Jeff earlier about the order of processing -- by accident, of course)...
Noise Reduction is applied before deinterlacing. This can also help the deinterlacer better detect what is moving (and not just noise). The NR I use in this script is very low-tech, but extremely fast and leaves very little artifacts when used in moderation.
Try NRStrength=15 (or, say, between 10 and 25). This also leaves you with video that will reveal less noise after color correction and that's more compressible for DV, MPEG2, etc.
EDIT: Very Important. I'm assuming the test clips you are using are at or about 1/60th of a second shutter speed. Correct? Faster speeds will look like "ghosting", but 1/60th should look very much like natural motion blur when using BlendMethod=4 or 5 (or any method, to some extent).
>it is best to isolate the deinterlacing and blending as much as possible for testing.
OK. Let's have at it then.
I do. Sounds good.
>I suspect this is the same issue that occurs in DVFilm Maker
I can't say it wasn't from the same cause, but it wasn't the same visually.
>I'm assuming the test clips you are using are at or about 1/60th of a second shutter speed. Correct?
As they should be. Even Frames and DVFilm Maker come with this caveat.
So, my first uploads of several to come. We start with the original unprocessed video and audio, minus the music track and opening credits to protect the innocent. It's a little strange without the music, but there are many good test scenes in this real world opening edit from a real client's wedding.
The clip was edited using the original DV captures in a Standard 30i project, encoded out of a Premiere to MPEG at 30i using CBR 7, and should be usable in Encore without further transcoding. This clip will act as a reference. The audio portion will not change as more clips come, so just the one file can be used for all of them.
> OK. Let's have at it then.
Rendering as we speak... It will take a while. It's a sort of cheesy fly-though animation. The rendering is long because I'm using camera motion blur to simulate 1/60th shutter speed on the 60fps -- and 1/24th + 1/48th on the 24p, respectively. As this test is primarily for HDTVs, it is 16x9 widescreen.
Something else I've thought of: 1/24 is a bit blurry, and most filmakers use 1/48th (I think). DVFilm Maker simulates this by alternating the blurred frame (as does BlendMethod=3). BlendMethod(s) 4+5 will create the 1/24th look from 1/60th DV sources. If the blur gets a bit heavy in these modes, try using bR=0.75 instead.
> I can't say it wasn't from the same cause, but it wasn't the same visually.
Yep. It may be just a contributing factor. Could you maybe compare/contast the types of artifacts you saw here vs. DVFilm Maker?
Oh my god... I was just watching your clip, Jim. I'm pretty good at lip reading and I swear that in that shot around 00:01:02 the bridesmaid says "I don't know what the f***...". That's great :)
>1/24 is a bit blurry, and most filmmakers use 1/48th (I think)
It depends on the film speed and shutter angle. 180 degree shutter is probably the most common for well lit scenes at 24 fps, so the shutter speed translates to 1/48th.
>Could you maybe compare/contrast
You can do that yourself, clips are coming.
You're right, Jim. Some of these shots are challenging to deinterlace. If you get a chance, could you please upload the original DV clip of the second shot? (the pan-down of the house)
It is interesting because it has thin, sharp, near-horizontal lines moving vertically, then becoming static. That's a toughy. Good test material for me.
I've been thinking also about sharpening techniques and yet another blending method... be prepared :)
>Good test material for me.
That's why I've been using it since last September as a 'problem clip'. The pan of the trees near the end is once scene as well. In fact, there are over a dozen shots in this sequence that give me problems of one sort or another, depending on the conversion method used.
The DV clip is being uploaded now.
Jeff, Jim... You may want to hold off on testing BlendMethod(s) 4 and 5: I found a little typo (a "1" instead of a "2") that makes the blending too weak. It'll be fixed and uploaded tonight.
BTW, Jim -- It looks like (at least) some of the flickering you saw with was due to the "DVFilm Maker-like" blending with Basic=true. It's gone with the other methods.
I am also strongly leaning towards DeintMethod=1 as the best option.
>You may want to hold off on testing BlendMethod(s) 4 and 5
Too late. BlendMethod=5 as implemented in the version I have makes the judder on pans and other cross-frame motion significantly worse. BlendMethod=4 blurs more (and therefore gives the appearance of more fluid motion), but the underlying judder is increased above BlendMethod=0 (the default).
Right now, the very best I have seen so far on my HDTV is dv60i2film(DeIntMethod=1, iM=2, bR=1.0, bA=2.5). This setting also comes in second on the CRT, just a fuzz behind Jim's previous favorite iM=0, bR=1.
My next question is:
If I want to smooth the faster motion a bit more, but not blur to the same level as BlendMethod=4, should I decrease bR towards 0.5 *and* increase bA toward 3.0 or 4.0? Or should I just increase bA?
Re-encoding in Squeeze using a GOP structure of N=12 and P=3 allows Encore to import the Squeeze files as "Don't Transcode", whereas N=15 and P=3 would not. Go figure.
Well, that GOP thing explains it somewhat. However, I was inadvertently using a GOP of 15 for my 24p exports out of Premiere, and they came into Encore just fine.
I tried DeintMethod=1, but "there is no function SecureBob".
You need MVBob() installed for DeintMethod 1,2 or 3:
Thank you, will try again.
I figured out a part of why I was having issues running scripts a while back, and this is data that may affect others.
I have my DV originals stored in the folder E:\Client. Inside that folder, I create two more folders, an AVS folder to put the batch scripts into, and a 24p folder to store the converted media. So, scripts coming from the Batch Scripter program get stored in E:\Client\AVS, and converted media will end up in E:Client\24p. Seems simple enough.
The problem is that when I ran the Batch Scripter program to test some different settings, it had created erroneous scripts, telling Vdub to look in the 24p folder for the originals which, of course, were not in there. No matter what I did, I could not get Batch Scripter to create scripts telling Vdub to look in the original Client folder. It kept incorrectly adding the 24p folder to the path name, so Vdub couldn't open them because it couldn't find them.
The solution was to temporarily move the 24p folder out of the Client folder, then create the scripts, then move the 24p folder back.
Seems like a bug with Batch Scripter.
> BlendMethod=5 as implemented in the version I have makes the judder on pans and other cross-frame motion significantly worse.
BlendMethod=5 should now be improved quite a bit, judder-wise. New version!
> BlendMethod=4 blurs more ... but the underlying judder is increased above BlendMethod=0
That's true. It may be the ideal option, however, for low motion scenes. There's just no way around it: fluidity (blur) vs. stagger (crisp), unless ...
> the very best I have seen so far on my HDTV is ... iM=2, bR=1.0, bA=2.5
... you use motion estimation! But this is very slow (and can have some strange artifacts sometimes). These settings are really quite perfect, however, at simulating 1/48th shutter speed. (120 / 2.5 = 48). It was always my hope the blending would hide the mo-comp anomalies. It usually does, from what I've seen.
Yes, BlendMethod(s) 4 + 5 are a bit soft, but can be a reasonable (fast) alternative to motion estimation. Try to lower bR (maybe to .85, .75, .65, etc.) until the wobble starts to offend you ??
1. BlendMethod=5 is not good. On either CRT or HDTV, it's pretty ugly. At its worst, the ghosts are visible and persistent enough to give the feeling that you need 3D glasses to watch.
2. BlendMethod=4 is pretty nice. It's blurred more than I would like, but the motion is smoother than other settings, including, maybe the 1/48th shutter speed setting. Low/no motion is good.
3. I'd like to get the best of both worlds: either smooth out the 1/48th shutter settings by allowing a touch more blending/blurring, or sharpen up BlendMethod=4 by toning down the blending/blurring a tiny bit.
I guess I should increase bA to 3.0 for the 1/48th shutter setting and/or decrease bR for BlendMethod=4 to 0.75? The reason I ask rather than just try it is to find out from Dan if the proposed corrections theoretically go far enough to be noticeable/effective.
What do you think?
Well, it seems I improved BlendMethod=5 in concept but not in deed. I found out exactly what Jeff is talking about (the 3D glasses effect) and beieve me, that was not my intent :) I'll fix it (again)...
Jeff and I had an email exchange in which he asked me a lot of very specific questions about how all the blending works. It forced me to think about it myself. I'll post a distillation of the descriptions I gave Jeff when I get my own head around all of it.
I just tried four new methods on both SD CRT and Sony Bravia High-Def LCD (interlaced component from DVD).
1. (Basic=true, DeintMethod=4, bR=1)
2. (iM=2, bR=1, bA=2.5)
3. (DeintMethod=1, iM=2, bR=1, bA=2.5) [Jeff's recommendation]
4. (DeintMethod=1, iM=0, bR=1) [slight mod of my old favorite]
The first looked pretty good on both displays, but did produce some jaggies, and it did produce some unacceptable flicker in trees on the HD set.
Two produced less jaggies, and less flicker, but did also have some added flicker on a pan of table legs near the end.
Three looked best overall on the HD set, and very, very nice on my SD set. It had fewer jaggies still, and even less flicker than one or two on the HD set, with no flicker on the SD set. The only issue was a really weird 'splotch' during a pan of a stairway balustrade. It reminded me of the kind of splotch you would get if a real frame of film were left static in front on the projector bulb for too long, but now imagine this was done for every single frame over two or three seconds. Each splotch would be slightly different, and the whole 'blob' would appear to shimmer when the frames are viewed full speed. Not sure what might cause that. It appeared on both SD and HD viewing, and was the only real artifact of concern on this test. Overall, I could definitely give this to a client and be satisfied it'll look good on whatever display they use to watch it.
Four didn't have the splotch, and had the best overall "lack of jaggies", but it did look soft during pans of trees, and even a little stuttery. One, two and three were sharp throughout the pans, and seemed more fluid, even more so than my old favorite of (iM=0, bR=1).
So far, three seemed the best option, despite the weird splotch. Unfortunately, it also took the longest to process, about 33% longer on average over the other three options, and almost 5 times longer than (iM=0, bR=1).
P.S. I tried the OutputPAL option last night for a client whose parents live in Poland. Took a while (18 hours for less than 90 minute video), but the results came out very, very nice. Well done, Dan.
Thanks, Jim. It is hard work, but I am obsessed now with making this the best it can be.
> (re:#1) ... pretty good on both displays, but did produce some jaggies, and it did produce some unacceptable flicker in trees on the HD set.
This is as much because of the blending as the deinterlacing: Like in DVFiln Maker every other frame @ 24p is a single, unblended field that will show every jaggy edges left behind from any deinterlacer (as none of them are perfect). Even without jaggies, BlendMethod=3 can flicker on subtle motion as it vacillates between unblended and blended frames.
> (re:#3) ... The only issue was a really weird 'splotch' during a pan of a stairway balustrade
Therein lies the peril of motion compensated framerate adjustment :) There can be weird artifacts that are "tweened" on the interpolated frames. It's not really something you can predict, but oddities do happen. When it works, however, it works fabulously. How big was the "blob"? Does it at all resemble any element of the original scene?
As for faster, non-mo-comp'ed blending modes, try BlendMethod=4. This is a little soft -- very much like 1/24th shutter speed -- but very smooth. Try gradually reducing bR until the wobble and jaggies become offensive, trying to tease-out the sharpness. Start at bR=0.95 ... bR=0.85 ... bR=0.75 ... bR=0.65 ... Try to find the "sweet spot".
You may also want to investigate (BlendMethod=6, bR=1). I haven't really tested it much myself, but it's a variation of Blendmethod=3 that should smooth out jaggies and wobble a little more, hopefully without adding too much to the overall appearance of blur.
> I tried the OutputPAL option last night
Let's say you've already made your NTSC Film 24p version and saved it to a (preferably lossless) file. Make a script like this...
ar = last.audiorate
... and save yourself much time :)
>Make a script like this...
Who IS this guy?
>I am obsessed now with making this the best it can be.
You and me both. I'm very encouraged by what we have so far.
>How big was the "blob"?
I'm uploading the original DV clip so you can see for yourself.
>Let's say you've already made your NTSC Film 24p version...
Good script to have for the future. In this case, the original was only a DV version due to the package ordered.
I've said it before, and I say it again...if Adobe doesn't consult Dan on CS4, we are all the poorer for it...these threads are very interesting - I've no idea about half of what they are talking about, but it's all interesting
Care to participate in the review process from Part I of the series? Go to post 199. You don't need to know anything technical, just download the clips, bring them into Encore and burn, then sit back and watch. Your opinions here are valuable.
Thanks, Jim. I'll let you know what I find. I am shoulder deep in editing projects right now, but I plan on starting in on all this as soon as I'm free.
I'm really hoping that CS4 is out soon and I won't need it (Adobe, have you been reading any of this?)...but it's nice to know all this information is here if we do.
Great work all!