[MPlayer-users] denoise3d and encoding DVD

[Corey Hickey]

D Richard Felker III wrote:
> [Automatic answer: RTFM (read DOCS, FAQ), also read DOCS/bugreports.html]
> On Fri, Feb 07, 2003 at 08:19:37AM -0800, Corey Hickey wrote:
> 
>>[Automatic answer: RTFM (read DOCS, FAQ), also read DOCS/bugreports.html]
>>I've found that most of my DVD's have a certain high-resolution
>>grainy-ness to them, and that by using the denoise3d filter like this:
>>-vop scale=700:376,denoise3d,crop=716:480:2:0
>>right after cropping and before scaling I can increase the PSNR
>>somewhat. I've thrown a few short clips into the test, and it seems for
>>high motion scenes the PSNR increases by about 0.5, whereas for low
>>motion scenes the PSNR is raised by up to 2.
> 
> 
> This is a bogus measurement. When lavc tells you the PSNR, the
> "signal" it's comparing the encoded picture to is whatever you feed
> into it, i.e. the already-filtered frames. Of course frames without as
> much noise in them are going to encode better, but they won't
> represent the original movie as well since some information was lost
> in the filtering. If you could measure the PSNR relative to the
> original picture, it would probably be worse with denoise3d.
> 
> On the other hand, that doesn't mean it's a bad idea to use denoise3d
> for this. When you reduce the noise (which is likely not to be
> perceptually important to a human viewer watching the movie), you'll
> conserve bits which can then be used for more accurately encoding the
> details you actually want to see.
> 
> My guess would be that denoise3d is probably a win in terms of
> improving visual quality, at least in most cases, but PSNR is not a
> good way of measuring whether or not it's a win. So if you really want
> to check, you need to do double-blind tests with a group of viewers
> and see which encode people identify as looking 'better'. Or else find
> a good mathematical model for perceptual quality (good luck! ;).
> 
> 
> Rich
> 

Ahh, of course, that makes perfect sense. I'll go ahead and run some
tests by my friends (can't get a statistically valid sample size - I'm
not _that_ popular; but, I'll see what I can do). Maybe when I have some
time I'll prepare some samples and post them to the list here.

Thanks for the tips,
Corey