Strange behaviour using Telecide in Decomb filter

[djloewen]

Hello all,

I am encoding a video from a DVD. I usually use Gordian Knot, but with Decomb 5.2.2 - today I also changed to AVISynth 2.5.7 and DG 1.4.9.

The source I am encoding is originally 24fps, hard-telecined, and I am trying to use Telecide on it. The script line I typically use for such a task is:

Telecide(guide=1,vthresh=20.0,blend=true).Decimate()

I remove GKnot's "order=" variable because it does not work with 5.2.2. My intention with the other two parameters is to match those of the FieldDeinterlace() command, so that I am effectively IVTC'ing and then deinterlacing to get rid of the mouth jaggies and such. This has worked very well for me in the past.

In this case, the final video seems to have about 40% mismatched (but deinterlaced/blended) frames. My first instinct was that perhaps because of the two other upgrades I did today, Decomb was no longer receiving the correct field order info (DGindex says BFF).

To test this, I began by running a script with the vthresh/blend parameters removed, and discovered to my shock that the video IVTC'd beautifully.

Just to double-check, I loaded up the scripts side-by-side in VirtualDubMod. They are standard GKnot scripts, with one line difference between them. I will attempt to post a screenshot of the results. On the left is:

Telecide(guide=1,vthresh=20.0,blend=true).Decimate()

on the right:

Telecide(guide=1).Decimate()

Now, my understanding of Telecide was that the vthresh and blend parameters were strictly post-processing, and therefore should not have any effect on which fields get matched together. What is going on here? What am I missing?

Thanks,
Dave

[Guest]

Post a link to an unprocessed source stream sample that can be used to duplicate your issue.

IIRC, vthresh can affect the matching only via the 'back' option.

I think the match is the same, but the difference you see is due to one using blend=true and one to blend=false (default when you don't specify it), i.e., blending versus interpolation.

[djloewen]

Thank you for the response. You are right, it was the same match: when I set post=0 to the second clip, there was clearly a problem. And yes, it turns out it was a field order problem. DGindex reported it as being BFF, but it is in fact TFF. So, AssumeTFF() will do the job just fine here, but... what might the cause be of DGindex misidentifying it? Might a bad rip cause it? Does it just happen occasionally? The reason I'm concerned is that I had never seen this before until I switched to DG 1.4.9, so I need to know whether I should switch back to 1.4.3. Any thoughts? I will happily upload a clip of unprocessed video (if someone knows of a good website, it would save me a google search), but I sadly have no idea how to trim off a small bit of a VOB/MPEG2 stream. VirtualDubMod doesn't seem to like direct-stream-copying on it....

Also, while I have your attention, perhaps you could answer something that I have long wondered about. The Decomb Reference Manual says "For NTSC 25fps->30fps telecine guidance, set guide=3". In my experience, most 25->30 material has four interlaced frames and two clean ones, but I have heard that this is not always the case. The Decomb Tutorial then mentions 3:2:3:2:2 for guide=3, which to me suggests two interlaced and four clean. So, does it look for one specific pattern, or is it good for any material which repeats one frame in six?

(Also, if you happen to have an answer, I would love to know why 25->30 material never seems to be a "clean" telecine job, but rather the level of interlace seems to drift larger and smaller over the cycle of six frames...)

[Guest]

DGindex reported it as being BFF, but it is in fact TFF. So, AssumeTFF() will do the job just fine here, but... what might the cause be of DGindex misidentifying it? I need to see the source stream to answer that.

I will happily upload a clip of unprocessed video (if someone knows of a good website, it would save me a google search), See here:

http://forum.doom9.org/showthread.php?t=96362

but I sadly have no idea how to trim off a small bit of a VOB/MPEG2 stream. VirtualDubMod doesn't seem to like direct-stream-copying on it.... Get my tool DGSplit. Then set the split size to (say) 10-20Mbytes and start the split. Stop it after DGSplit starts writing the second file. Then upload the first one.

So, does it look for one specific pattern, or is it good for any material which repeats one frame in six? I can't remember what I implemented. :) I'll have to check that and get back to you.

(Also, if you happen to have an answer, I would love to know why 25->30 material never seems to be a "clean" telecine job, but rather the level of interlace seems to drift larger and smaller over the cycle of six frames...) I don't know what you mean by "level of interlace".

[djloewen]

I have uploaded two sample clips:
http://rapidshare.com/files/115122246/Sample1.VOB.html
http://rapidshare.com/files/115122247/Sample2.VOB.html

The first is a sample of the clip that is being incorrectly reported as BFF by DGIndex. Help!

The second is an example of a 25->30 conversion that exhibits the "drift" I was talking about. When stepping through the unprocessed frames, the interlacing appears to move in a sort of circular pattern, from one side of the image to the other, and then back to center. This probably sounds ridiculous, but I hope the clip will help. When stepping through the individual fields, I now see what appear to be a number of blended fields. A fascinating side-effect of this is that when I Telecide with guide=3,post=0 (no decimate), there are indeed two similar frames in every cycle of six, but that the first of these two frames appears to be interlaced on the left side of moving parts, and the second appears to be interlaced on the right side. Of course, with decimate and post (with blend) it looks great. I have seen this pattern (likely with blended fields) in every 25->30 conversion I have ever come across. Do you know if this is standard practice, or what? With the 3:2 pulldown it's so easy to get a nice 24p picture out of it, it would be nice to do the same with 25p. Especially since ATSC is the new thing and everything will be able to play 25p natively anyway...

[djloewen]

lol sorry, I picked the first one on the list :)
Here's Sample2.VOB:
http://www.megaupload.com/?d=PX95A5N7

[Guest]

The first sample exposes a deficiency in DGIndex that I will repair for 1.5.0 final. When the clip is field structured, the TFF flag must be 0 and does not signal the field order. The field order is obtained by seeing which field picture appears first in the bitstream. Almost all field structured clips are TFF.

The second one is field blended as you say and I have a well-known aversion to dealing with that stuff. You are on your own. :) In any case, it raises no issues for DGIndex.