[XviD-devel] Fades

[Christoph Lampert]

On Mon, 18 Nov 2002, daniel smith wrote:

> i worked a while back on adapting ME to fully match epzs.  during playing with subsets i figured why not add a "subset d" which is just the 0,0 vector checked with a mean-removed SAD, only to be performed if subsets a,b,c all fail to meet a threshold.
> 
> it wouldn't catch every fade, but would certainly help (and not be all 
> that expensive).  might even help with some strobe effects.

I guess normal SAD won't catch even the simplest fades, because there's a
bias in luminance between current and reference. But we could use mrSAD
for this last check, if we find some constants to compare SAD-values to
mrSAD-values. Somehow I'd guess both would be very comparable, because
only DC coefficient is missing, but my old tests showed that mrSAD values
are much lower. 
Range of values for SAD and mrSAD should be comparable, because
otherwise we'd have to modify mode decision etc.

Is anyone willing to test this? I'll be too busy with my PhD for the next
4 weeks, but I just did a 2CD Rip of "Forrest Gump" and image was perfect
except for the very first and the very last 5 seconds, which were simple
fade and so extremely blocky, I had to cut them out :-(

gruel