[XviD-devel] GME wth random sampling

[Mårten Björkman]

Stupid Micro$soft! Why don't they ever do what you expect them to do? I 
repeat!!

I ran a series of tests with random sampling in GlobalMotionEst() in order 
to suppress outliers in the motion vectors. The sequence I used was Mobile 
Calender, since the motions in this sequence are planar, they should be 
relatively easy to capture

Obviously, using random sampling the estimated global motion better matches 
those of the motion vectors. The number of skipped blocks went up from 20K 
to 94K and after refinement from 56K to 130K, which means more than 40% of 
all blocks. This change led to a much larger number of S-VOPs.

However, the coding did not improve. The size of the coded vectors went down 
quite a lot, but the textures generated more bits. Depending on the settings 
it sometimes improved the coding and sometimes not.

Why do the textures generate that many bits? The numbers of intra blocks 
stay approximately the same, so the extra bits come from the inter blocks. 
Could it be that errors are introduced in the warping and these errors 
aren't modelled properly?

_________________________________________________________________
MSN 8 with e-mail virus protection service: 2 months FREE* 
http://join.msn.com/?page=features/virus