[x264-devel] Re: [PATCH] PSNR averages

[Laurent Aimar]

On Sun, Aug 08, 2004, Måns Rullgård wrote:
> Laurent Aimar <fenrir at via.ecp.fr> writes:
> 
> > On Sun, Aug 08, 2004, Måns Rullgård wrote:
> >> Loren Merritt <lorenm at u.washington.edu> writes:
> >> 
> >> > x264 takes the arithmetic mean of frame-wise PSNRs to produce an
> >> > overall PSNR. This is meaningless. (Extreme case: what do you do with a
> >> > frame with no error? (infinite PSNR))
> >> > This patch corrects it to calculate PSNR from the mean squared error.
> >> >
> >> > I also replace "PSNR/Size" (which is similarly meaningless) with
> >> > "MSE*Size". This quantity is probably closer to what someone wants to
> >> > measure when they think of "PSNR/Size". It also tends to be relatively
> >> > constant for a given video over a wide range of QPs.)
> >> 
> >> Speaking of PSNRs, how much is the overhead of calculating these?  If
> >> I only want encoding speed, can anything be gained by disabling some
> >> calculations?
> >  In encoder/encoder.c, I added a #define DEBUG_PRINT_STAT
> > comment it out, and it won't print any stats.
> 
> I was thinking on a lower level.  To calculate the PSNR, the quantized
> samples must be dequantized and inverse transformed, as is done in
> various places.  Is the reconstructed image needed for other reasons
> than statistics?  Otherwise I suppose some cycles could be gained by
> skipping these steps.
 The quatif/dequant has to be done anyway (as the encoder need the
reconstructed picture to do the MC). Not doing PSNR will only save a SSE
over the complete reconstructed picture.

-- 
fenrir

-- 
This is the x264-devel mailing-list
If you are in trouble, please contact <postmaster at videolan.org>