SysKin's latest ME changes

[octapus]

Anyone tested the new Motion Estimation which Syskin posted at 25/9?
These changes exist only in latest Umaniac's build if i am not mistaken.
If i got it right, they are supposed to make Hint MV file give better quality and faster.

[Koepi]

They are broken at the moment.

[sysKin]

Originally posted by Koepi
They are broken at the moment.

Yes, but only with lumimasking. Hinted ME withoul lumimasking works (for me -> doesn't mean anything).

[-h]

Yes, but only with lumimasking. Hinted ME withoul lumimasking works (for me -> doesn't mean anything).

I remember having to do some "ugly" things to get lumi masking to work with hinted ME, like forcing blocks to intra if lumi masking tried to set dquant for an inter4v block. It was just better in the end to enable lumi masking in both passes.

What exactly are you doing above and beyond the hinted MV? Just a diamond search, or an additional inter4v/inter/intra check?

-h

[sysKin]

Originally posted by -h
What exactly are you doing above and beyond the hinted MV? Just a diamond search, or an additional inter4v/inter/intra check?


Well it's quite simple, really:
- if hinted data says that MB was intra or not coded, continue;
- do a check of hinted vector
- do 3 more checks of 3 more hinted vectors, if MB was inter4v
- do a diamond starting with the best found so far
- do halfpel refinement
- do inter4v search (if allowed) - in the best position found so far (diamond + halfpel)
- decide inter/inter4v (if former is allowed). The end.

Inter4v is not allowed if it's turned off in configuration or if quantizer changed.

Anyway, the bug is there even if inter4v is turned off in both passes... It looks like a bitstream problem - there are pink macroblocks, wrong vectors, random DCT data.....
I'll look into it, of course.

Radek