[XviD-devel] Qpel Problem Samples

[Michael Militzer]

Hi,

Quoting Christoph Lampert <chl at math.uni-bonn.de>:

> thanks for the samples and detailed description. We'll check... So far we
> know that _if_ Qpel is helpful, it already better for any short parts
> of the video, and usually not by higher PSNR, but by lower filesize. 
>
> So if you want to decide on QPel/noQPel you can encode a
> short piece (10s) from the middle with fixed quant and then decide on
> QPel/noQpel by smaller filesize. 

I don't agree. You can't decide from just a short test sequence if qpel
would be helpful or not for the overall clip. With 'Saving Private Ryan'
for example, there are some scenes were qpel helps greatly and others
where it does not help at all. I haven't checked yet if qpel is a overall
gain or not for this movie...

> Maybe somebody could add such a feature the GUI? An "Auto" setting
> for QPel?  Because MPEG-4 doesn't permit to just switch QPel on or off
> within a clip.

yes, unfortunately qpel can't be adaptively switched on or off within a
clip. I checked the sample clips that had been posted over at doom9
(thanks for that btw), and qpel indeed performs rather bad on these. How-
ever it seems that b-frames don't seem to work great either on these
clips (but I have to recheck). What should be also figured out is whether
the qpel encoded clips actually look worse _visually_ than the halfpel
encoded clips. It seems that qpel introduces or amplifies noise which
leads to a more 'crisp' image (similar to MPEG vs. h.263 quantization).
This additional noise will lower PSNR but it doesn't have to lower the
visual 'watching experience'...

bye,
Michael

PS: I uploaded some new XVID_ME options yesterday that make qpel quite a
bit faster. You might want to check these out. If we can't increase qpel
quality, we can at least try to reduce its computational overhead ;-)


> On Thu, 4 Sep 2003 j_coleman at shaw.ca wrote:
> > It was mentioned on the doom9 forums by Isibaar that if we came across some
> 
> > samples where qpel performed worse than half pixel (ie lower PSNR) that we
> 
> > could put them online and post a link.  I posted this on doom9 but I don't
> 
> > know if any of you guys have seen it yet.  So anyway here are the source 
> > files that I have found that when encoded with qpel exhibit a lower PSNR. 
> 
> > Probably though I should say something about how they were encoded first. 
> 
> > All three files were force-filmed to 23.976fps in dvd2avi and this avs
> script 
> > was used or something similar: 
> > 
> > LoadPlugin("C:\Program Files\Avisynth 2.5\plugins\Mpeg2dec3.dll")
> > Mpeg2Source("C:\TERMSE_SIDEA\VIDEO_TS\terminatortest.d2v")
> > 
> > They were then loaded into virtualdubmod and encoded with a dev-api-4 build
> 
> > from August 19th built with ICL7.0 just a win32 release, only default
> options 
> > were used except of course for qpel.  All encodes were two pass and PSNR 
> > measurements were performed with psnr4avi.  Anyways here are the results:
> > 
> > For a short clip aiming for 2MB:
> > with qpel: Average PSNR (Y U V): 43.26, 45.70, 45.76 Average PSNR: 44.91
> > without qpel: Average PSNR (Y U V): 44.18, 45.86, 45.96 Average PSNR:
> 45.33
> > And here is the url for the source file
> http://www.geocities.com/bonzi5252/..
> > .atorsample2.zip
> > 
> > For another short clip again aiming for 2MB:
> > with qpel: Average PSNR (Y U V): 44.15, 48.54, 49.01 Average PSNR: 47.23
> > without qpel: Average PSNR (Y U V): 45.15, 48.67, 49.13 Average PSNR:
> 47.65
> > Source file can be found here
> http://www.geocities.com/bonzi5252/bondsample.
> > zip
> > 
> > And one more clip aiming for 350KB:
> > With qpel: Average PSNR (Y U V): 45.26, 48.94, 49.09 Total Average PSNR:
> 47.
> > 76
> > without qpel: Average PSNR (Y U V): 45.94, 48.98, 49.14 Total Average PSNR:
> 
> > 48.02
> > And the source file http://www.geocities.com/bonzi5252/bondsample2.zip
> > 
> > If you need more samples or if these ones are completely useless, don't 
> > hesitate to ask or comment.
> > 
> > _______________________________________________
> > XviD-devel mailing list
> > XviD-devel at xvid.org
> > http://list.xvid.org/mailman/listinfo/xvid-devel
> > 
> 
> _______________________________________________
> XviD-devel mailing list
> XviD-devel at xvid.org
> http://list.xvid.org/mailman/listinfo/xvid-devel
>