[MEncoder-users] on psy_rd

[Grozdan]

2009/9/18 Matyas <mplayer.list at sustik.com>:
>
>
> I think I understand now.  In order to have this "similar complexity" the
> (total) bitrate is upped.  I thought that psy_rd does reallocate bits to
> preserve detail (complexity?) where it is considered more important for the
> human perception.

psy-rd attempts to maintaining detail/complexity. Since more
detail/complexity requires more bits and since psy-rd biases towards
such detail/complexity preservation, the bitrate will be increased. It
also tries to preserve as much as close to the original, including
grain/noise, and so again this results in some bitrate increases (note
that psy-rd may introduce a bit of ringing, same for psy-trellis).
There are "special cases" where psy-rd could potentially make parts
worse - for example, where in a grainy picture with high contrast,
bits are moved from the front to the back in order to maintain the
grain, thus decreasing quality somewhat in the front of the picture,
where, for example, a person is standing.... In the end, both psy-rd
and adaptive quantization (aq) redistribute bitrate to places where it
matters, for aq it tries to minimize blocking in flat areas/textures
thus it increases bitrate in such places, where as psy-rd tries to
preserve the detail/complexity of the image. If you want to test, best
is done so with very low bitrates and then compare encoded without
psy-rd/aq and with psy-rd/aq

>
> I read some forum posts and comments, which made me believe that psy_rd
> contributes to a compression improvement.  Meaning that using it will either
> allow you to use fewer bits to get the same quality (as human perception
> goes) or better quality with the same bitrate.  So it appears the parm
> controls (affects) how many bits are used without the promise of a better
> compression.

psy-rd contributes to visual improvement and since it is more
important than staring at some numbers which hardly express it
accurately, I think you should test various strengths and see what
looks good to you. Not everyone agrees on what's "best" ;)

>
> I also run the experiments comparing psy_rd=0.0:0.0 to psy_rd=1.0:0.0.
> Again, the second run featured the higher bitrate.
>
> My confusion came from that I expected the encodings using psy_rd to have
> *lower* bitrates.  (I was also suspicious when I got a quite large bitrate
> impact from a quite small change: 0.0:0.0 -> 0.0:0.2.  But that can be
> chalked up to my inexperience of the scaling of this particular parm value.)

Psy-trellis is still considered experimental so it could behave somewhat "weird"

>
> I have encoded some OTA 1080i sources.  (Deinterlaced, denoised and scaled to
> 960:540 first and encoded with qp=0, then run the rest of the suite.)  The
> original has some visible compression artifacts (broadcasing people if you
> are listening, you need to update your hardware/software!!!).  I could not
> tell the difference between crf=24 and crf=23.  (All artifacts I noticed were
> in the original as well.)
>
> So using 20 seems quite low to me.  Maybe crf values for larger pixelcount
> should be naturally higher?  (I do not have SD sources encoded, I will do
> some tests.)

I haven't done any HD encodes yet so can't comment on which CRF value
to use for them but I read on more than one places where generally a
higher CRF is used for HD and a lower one for SD. I use 20 for SD but
am thinking in going to 19 since most of the time for a picture of
720x352 (or a bit higher on the height side) I get a bit too low
bitrates for my taste, somewhere between 600-700 kbps but I try to aim
for 800-1000 kbps

>
> Have a good day,
> Matyas
> -
> Every hardware eventually breaks.  Every software eventually works.
>
> --
> Matyas
> -
> Every hardware eventually breaks.  Every software eventually works.
> _______________________________________________
> MEncoder-users mailing list
> MEncoder-users at mplayerhq.hu
> https://lists.mplayerhq.hu/mailman/listinfo/mencoder-users
>
>