[Ffmpeg-devel] [PATCH] DV image quality improvement

[Christian Iversen]

On Saturday 25 February 2006 23:18, Rich Felker wrote:
> On Sat, Feb 25, 2006 at 09:59:32PM +0100, Christian Iversen wrote:
> > On Saturday 25 February 2006 21:18, Dan Maas wrote:
> > > The attached (very small) patch improves image quality from ffmpeg's
> > > DV encoder, increasing PSNR by about 3dB on the BBC's test footage.
> >
> > Not at all commenting on the quality of the patch in question (which is
> > probably great), when are "we" (meaning "you" :-) going to stop using
> > PSNR as a quality measure? There is only a quite weak correlation between
> > percieved image quality and PSNR.
> >
> > There are other quality measures available that will give much better
> > results. A difference of 3dB really proves very, very little about how
> > good the result looks.
>
> Umm, as much as PSNR sucks, 3dB is HUGE. In ffmpeg development, 0.1dB
> improvement is even considered fairly good, provided there are no
> visual problems. 

I misremembered the Scale of Hugeness for PSNR. Mea Culpa :-)

> I agree that PSNR in general doesn't say much about video quality, but once
> you fix your testing to a particular codec (and thus a particular type of
> quantization/error) it's fairly well correlated with overall quality, at
> least for single frames. 

Well, I don't agree completely, but you probably have more experience in the 
relevant field.

Having read the paper on the Universal Image Quality Index (did I get that 
name wrong?), I've seen examples where images there were very clearly 
completely different in quality got the exact same PSNR rating. I think it's 
safe to say that the author chose some examples where PSNR didn't exactly 
come out top of the class, but it still goes to show that things CAN go 
terrible wrong with it. Assuming it's "fairly well correlated" with image 
quality because of some fixedness in the situation is dangerous and 
unfounded, IMHO. But again, that IS just MHO :-)

> Often these 0.1-0.2dB improvements come with a perceptible visual
> improvement. Rate control makes the whole topic much more complicated of
> course, but that's mostly irrelevant with DV which is normally CBR. 

Well, I'd say that just the fact that there is movement at all makes the 
situation more complicated. Rate control adds on top of that.

-- 
Regards,
Christian Iversen