XviD + internal smoother filter

[Koepi]

Ahoy,

I asked the same question over at the videocoding.de forum in the general development area, but just to make sure it gets read I'll ask here again ;) (uhoh, isn't that cross posting as well? ;) )

In the current stuff I'm encoding there's a lot of noise when there's movement in the picture (I use temporal smoother to filter out most of the noise from the analog TV source). With nandub nn involved something like motion dependend smoothering - the DivX3 codec has an internal smoother filter (called "crispness" in the vanilla codec setup). This would be very useful for XviD as well.

Can you do such a thing as well? -h? nic? :)

Best Regards,
Koepi

[-h]

Isibaar is working on this area of pre-processing (motion masking in this case) - you might want to prod him for more details.

There was a bit of discussion on a much more complex masking method than MP43's unsharp mask filter..

-h

[Koepi]

You do not mean that the luma masking part is motion dependant?

Because when I tried unfiltered encoding the result was really bad (well, since that noise is counted as detail it's still very good, but maybe a little filtering would help it ;) )

Regards,
Koepi

[Teegedeck]

Hello Koepi,

lumi-masking isn't motion-dependent.

I was always pushing for that motion-masking thing, too. ;) Was one of the best features of DivX3.11.

BTW, what quantizer do you use? (For heavy noise, MPEG-quantizer just isn't a good choice.)

[Koepi]

I don't use quantizers, I use 2 pass (internal) shaping. Since i wanna do a 1 CD version I use what is advised everywhere:

1st pass, h263 quantizer scheme, no luma mask, 2nd pass with luma mask.

Works like charme, average quantizer is about 3.

The result is ok, with internal smoother algo it would look nicer I'd say...

Regards,
Koepi

[-h]

Quant=3 for 1 CD, without lumi masking?? That's one compressible source ;)

I'd go for MPEG quantization + lumi masking myself, but I'm not really a discerning viewer.

-h

[Koepi]

Second pass is WITH luma masking...

I filtered the source with a temporal denoiser, applied a bicubic resize (-0,75) to 640x352 (the TV picture is cropped 640x358....) - ah, and before all of that i deinterlace with blend fields. So the picture looks very nice after that procedure. The encoding doesn't :-/ Well, at least not as good as I expected - due to the lack of such (motion dependand) internal smoother.

That's how I found out about that ;)

Could anyone advise me to encode the movie with mpeg-quantizer? I use the DivX4 DS Filter for watching and the there should be some green artifacts, dunno how bad they are?
maybe the mpeg quantizer will give me the "brilliance" of the picture that I expect ;)

Well, if someone could implement a motion dependand smoother in a short time it would be better of course ;)

Regards,
Koepi

[Koepi]

I try another encoding now, with frameserving and processing through avisynth. Gives a little speed improvement and I don't have to leave the yuv2 color space... but there's no logo remover filter I know of ;)
This time I resize to 576x320 (*sniff* wherefore did I capture that movie with 640x480??).

Maybe the result is better now, the quantizer/DRF choice seems to be identical as if the size were 640x352...

Well, let's see ;)

Isn't there some pre-processing code for testing available? I'd love to "beta test" it ;)

Regards,
Koepi

[-h]

Oops missed the second pass note..

The MPEG errors are quite noticable once you know what to look for.. so just avoid the screenshots that've been floating around ;)

If you want, I could mock up a motion-masker that just decimates the high-frequency DCT coefficients if the block is moving, say, more than 10% of the image's dimensions. That's about the simplest pre-filter that I could think of (perhaps SAMPEG has some useful implementation) but I've no idea how useful it'd be.

Nor do I have any idea when I could mock it up :)

I'm finding that I want more and more to take a day off work..

-h

[Koepi]

Please, please, pretty please implement such a thing :) Even if it's just simple, maybe it can help. Since the codec is still very alpha/beta, it's good to test such things IMHO.

After my last exams for this semester tomorrow I'd like to give you a helping hand, I'll download the code then :)

Best Regards,
Koepi