Size DOES matter - DV related

[ulfschack]

I've seen mentions of it before on the forum. My movie doesn't fill out the entire CD and so on. This is often countered with that Div5 knows what it's doing and simply doesnt need more bits, and that it would be stupid to pad with junk bytes just to up the size.

Baloney I tells ya! I've tried every setting with my premiere frameserved home-DV. Psycho, qpel, b-frame, GMC, prefiltering, interlaced/non-interlaced source, bitrate, quantiziser 2-2..2-16. Only thing I've kept is fixed resolution, which is 512x384, and that's the way it's gonna stay.

Video comes out like crap and I can't get 35 minutes to fill more than 300 MB (roughly 1200 kbps). With that kind of shaky, panning material I ususally need double that with other codecs. Come to think of it, even my SVDCs (2600 mpeg2 interlaced) looks better on the computer than Div5. Remember now, I'm talking DV, a whole different ball game.

Now, I'm not looking primerely for a solution (wouldn't hurt, tho), but rather to warn other DV users about this wierdness. As of yesterday I'm back to 4.12 which is supieror under these circumstances. It actually gets <i>better</i> with higher bitrate!

Say what ?

[-h]

If you're fixing quantizers at 2-2, you're "maxing out" DivX5's quality. However, if you're using B-frames, the quantizer for said frames is always doubled. This could be the cause of your undersizing..

I've found DivX 4.x and DivX 5's 1-pass fixed-quantizer performance (with B-frames off) to be practically identical.

-h

[ulfschack]

I also tried without B-frames. What I forgot to mention is that all tests were with 2-pass, but it's interesting that you should mention similar behaviour for 1-pass (non-quality, right?), which means there's still hope :)

cheers

[RFeaster]

I've successfully encoded using a minQ of 1 using Divx5. This may increase your bitrate.

[theReal]

I have encoded one hour of PAL DV Video (uncut vacation movie) yesterday to compare it with a Gknot/Nandub encode of the same material.

I used a blend-deinterlace and precise bilinear resize to 640x464 in VirtualDub, which gave it the best compressibility with this resolution (I also tried resizing and/or filtering with avisynth - believe me or not, the compressibility was still worse even with soft bicubic resize and temporalsoften(2,3,3)...)

First, I made some compressibility tests using 2000 frames with Divx 4.12 to test the different filter/resizing setups at a fixed quantizer of 2. The best setup (VDub deinterlace, prec.bil. resize) reached 49MB. Then I encoded the same in DivX5 using b-frames and it was 32,5MB at the same or even slightly better visual quality.

Based on this result, I decided I should be able to encode the DV material at about 1/3 less bitrate than when I encoded without b-frames. I already had a GKnot/Nandub encode of the whole DV tape for which I had used a bitrate of 3000. So I decided a 2-pass bitrate of 2100 (i.e. about 45min of video on one cdr) might be enough for divx5 and gave it a try (using b-frames, no q-pel, no gmc, no psy).
The result is amazing: the visual quality is absolutely the same for the most parts and in some parts even better than the SBC encode (i.e. there is much less ringing around objects in front of a blue sky, the colors are brighter).

For me, this is the first divx5 encode that actually worked like it should, I guess because now I know I must not use widths not divisible by 16 and I should only use b-frames and none of the other new, fancy features (that often cause problems).
However, I have not encoded and muxed audio yet, but that should be the least problem :-)

So far for my experiences with Divx5 and DV Homevideo - ulfschack, I don't know why your DV encodes look crappy with Divx5, maybe you should just leave the default encoder settings as they are and try a 2-pass encode?
I also don't think there are any major differences between divx4 and divx5 as long as you don't use b-frames and all the other new features.
From what I've seen, I can't imagine Divx5 ever resulting in less quality than Divx 4.12

[ulfschack]

Thanks for for the tips and the extensive answer.

MinQ of 1 ? Is that some sort of hack? I must admit I never tried writing in a value of one ...

TheReal:
I believe that I tried all settings including the defaults, but without much impact on quality or size. Comparing it to the original DV tape reveals a big difference. In Div5 they've even increased the maximum bitrate to 10000 implying that now there's even more quality to get if your willing to pay the bits ... I see signs of the opposite. In my opinion it's now more geared towards full length feature films than ever, which is what most people want ... but not I.

On the other hand your figures show bitrates of well above 3000 even with the B-frames checked. And if you're able to do it so should I be. It migth be that the sequence I chose for testing was the worst possible, a weak link in the codec if you will. So ok, one last trip back to the drawing board before I kick this mo-fo forever :)

cheers

[Chibi Jasmin]

Originally posted by -h
If you're fixing quantizers at 2-2, you're "maxing out" DivX5's quality. However, if you're using B-frames, the quantizer for said frames is always doubled. This could be the cause of your undersizing..

I've found DivX 4.x and DivX 5's 1-pass fixed-quantizer performance (with B-frames off) to be practically identical.

-h

What do you mean quantizer is doubled for b-frames?

If I fix quantizers at 2, all b-frames have quantizer 4?

[-h]

What do you mean quantizer is doubled for b-frames?

The quantizer for B-frames is twice that of I- or P-frames.

If I fix quantizers at 2, all b-frames have quantizer 4?

Yes.

-h

[theReal]

Comparing it to the original DV tape reveals a big difference.You said you are resizing to 512x384, so of course the difference is big. This is what I found out as well: do not resize DV Video to less than 640x480 (better 640x464, it's closer to 4:3 according to Gknot).

I also tried converting DV to mpeg2 with tmpeg - with a a variable bitrate setting of 8000 it looked pretty nice, for very high bitrates mpeg2 has always been better than DivX. So if size does not matter, I'd recommend a very high bitrate mpeg2 encoding, preferably with no resize. If you want it any smaller (like < 4000), DivX is the better solution.

As I said before, I wouldn't go below 640x464 resolution for DV video - I tried 512x384 with different resizing filters and it looked bad no matter how I did it (btw. if your final divx is too small for your purposes, increasing the resolution will give you bigger files instantly!)

[ulfschack]

It's the old problem of getting it to play on grannies computer etc. It takes a helluva lotta CPU juice to run above 512x384. By that I mean above 500 MHz. Actually that's for 3.11 alpha that I started out with and it might be more optimized for that later CODECs, alltough I don't see why it should.

As for the difference in quality to DV. I've done this before. Twenty something DV productions to date, so I know what to look at. Both DivX and SCVD for backup and compatability reasons. 8000 kbps would give me ... what, 13 minutes on a CD if the standalone would even play it, which it wont. I realise you're only trying to help.

Previously I've been satisfied with 2300 kbps for divx which means that I can use the SVCD project and still fit it on a 80 min CD. Now I was startled to see that the quality was so bad for Div5. However, it turned out to be a particulary "bad" scene from the tape that I'd chosen to experiment on. In retrospect it might be as good as Div4 even if it only takes up half the space. It's just hard to believe such a vast improvement spacewise. Then again I suppose grannie's computer would have a hard time coping with the B-frames as well

I'm also curious about what kind of high definition material they had in mind when they changed the max from 6000 to 10000.

cheers