[MEncoder-users] Noise reduction and interlacing: (hq)denoise3d vs. nr

[Ivo]

On Monday 02 July 2007 23:40, Nico Sabbi wrote:
> Ivo wrote:
> >IMHO tfields/tinterlace is prefered if you run a temporal denoiser. If
> > you only (de)interleave as Nico suggested, the denoiser works with
> > different data and will perform worse.
> >
> >For example: (F=frame, f=field)
> >
> >Input F1(f1,f2) F2(f3,f4) F3(f5,f6) etc...
> >
> >Fictional denoiser uses current frame, and the next two frames.
> >
> >If you use il=d and il=i:
> >    F1 is denoised by using f1<-->f3<-->f5 and f2<-->f4<-->f6
> >If you use tfields and tinterlace:
> >    F1 is denoised by using f1<-->f2<-->f3 and f2<-->f3<-->f4
>
> that means that frame F1 is denoised using:
> - as top field a rework of  topf1<-->bottomf2<-->topf3
> - as bottom field a rework of bottomf2<-->topf3<-->bottomf4,
> right?
> I don't know the theory and I don't understand why a denoiser should
> perform better when dealing with fields generated by an interpolation of
> the actual and the other field(s)
> Can you explain, please?

Hmm, I must confess I hadn't taken top/bottom into account. The idea is that 
with tfields/tinterlace the denoiser does not skip fields and not look 
twice as far into the future. But the top/bottom shift complicates things. 
The look-ahead function of the above mentioned fictional denoiser will 
wiggle up and down in space as it progresses through time. il=d/i does not 
have that problem, but it looks twice as far.

Another thing, if the denoiser also does spatial denoising and you use 
il=d/i, it will blur part of the top of the bottom field into the bottom of 
top field and vice versa. I don't know if that's (very) visible but it 
might be ugly, depending on the radius of the blur and the dimensions of 
the fields.

I guess both methods have their drawbacks and interlaced content just 
sucks :)

--Ivo