[XviD-devel] I haven't been here for 4 days and what I see? :)

[Radek 'sysKin' Czyz]

Hi everyone,
I've been off-line for this weekend - now I'm back and I see GMC :-)
This is great.
Now, it's my turn to give you some news. I did more experiments
recently, and I think I should say that as much as TWO experiments
were a success. More or less.

1. I had some fun testing a different approach for adding motion vector
lenght to SAD in motion search. My reasoning was as follows: "if you
can find a good match, go for it (-> don't take d_mv_bits too
seriusly); if you cannot find a good match, then at least try to limit
number of bits for a motion vector (-> d_mv_bits more important)".
This may sound strange, but believe me, I used to test more crazy
ideas ;)
Anyway, my tests show that I can get about 0.15 .. 0.20 dB PSNR
improvement for free. Nice isn't it?
The only probem I'm currently having is inter4v search/decision - it
just doesn't work (total gain from inter4v below 0.05dB) and I have to
work a bit more on this.

2. Using chroma information for motion search. Everyone knows that
using chroma planes for motion search is slow, pointless an not worth
the effort. However, I wounldn't be me if I didn't check it myself,
right?
So I added chroma SAD to ME. The results are:
 - encoding is 5..7 % slower
 - PSNR gain for normal movies is 0.05.. 0.10 dB
 - PSNR gain for animation and computer graphics (tested on Warcraft III
 trailer) is 0.20dB and higher in high-motion&lots of colours.
 - "visual quality" seems to be better - there is no big and ugly
 colour blocks in the video (especially fast and colourful video)

It's possible that chroma info can help in noisy videos, and it may
also help with b-frames (simply because victors are closer to 'real'
motion).
The code is quite short and simple, it also works with qpel.
What do you think? Should I add a motion flag and implement it?

Well, I'll start re-implementing all the changes to new motion_est.c.
I'm waiting for your comments and posssible objections.

 Radek