[XviD-devel] [RFC] quality levels

[Marc FD]

This is a multi-part message in MIME format.

------=_NextPart_000_0023_01C23F90.789A16C0
Content-Type: text/plain;
	charset="iso-8859-1"
Content-Transfer-Encoding: 7bit

> Comments : this one seems to be more logical.
> 
> Why not  pushing these ADVANCED_DIAMONDS16/8  we talked in  a previous
> thread ?  (nb : already done in vfw).
> 
> So i'm  waiting for coments and  let's decide a common  quality set of
> flags.

Hi. i made a test on the XviD ME yesterday. take a look (attached)


------=_NextPart_000_0023_01C23F90.789A16C0
Content-Type: text/plain;
	name="XviD ME tests.txt"
Content-Transfer-Encoding: quoted-printable
Content-Disposition: attachment;
	filename="XviD ME tests.txt"

XviD Motion Estimation test with CVS 07.08.2002 (custom debug builds)
ME used :

1 ) EPZS_16 EPZS_8 : full EPZS(^2) (EPZS is not a good choise for =
search8)
PMV_EARLYSTOP16 | PMV_HALFPELREFINE16 | PMV_EXTSEARCH16 | =
PMV_USESQUARES16 |=20
PMV_EARLYSTOP8 | PMV_HALFPELREFINE8 | PMV_HALFPELDIAMOND8 |

2 ) EPSZ_16 PMVfast_8 : Best EPZS(^2) (mix between EPZS and PMVfast)
PMV_EARLYSTOP16 | PMV_HALFPELREFINE16 | PMV_EXTSEARCH16 | =
PMV_USESQUARES16 |
PMV_EARLYSTOP8 | PMV_HALFPELREFINE8 | PMV_HALFPELDIAMOND8

3 ) PMVfast_16 PMVfast_8 : PMVFast^2 (current Q6 ME in Koepi's build)
PMV_EARLYSTOP16 | PMV_HALFPELREFINE16 | PMV_EXTSEARCH16 | =
PMV_USESQUARES16 |
PMV_EARLYSTOP8 | PMV_HALFPELREFINE8 | PMV_HALFPELDIAMOND8

4 ) PMVfast_16 PMVfast_8 : PMVfast AdvDiamond (Advanced Diamond)
PMV_EARLYSTOP16 | PMV_HALFPELREFINE16 | PMV_EXTSEARCH16 | =
PMV_ADVANCEDDIAMOND16 |
PMV_EARLYSTOP8 | PMV_HALFPELREFINE8 | PMV_HALFPELDIAMOND8

Tested on anime content ONLY.(ME is critical on amime)
test clips :
1) Full Motion (a very fast scene with motion on the entire frame) (83 =
frames)
2) Moves (some scenes with motion on a still background) (536 frames)
3) Travellings/Zooms (551 frames)

All parameters others than ME to defaults
source : MPEG2 / VDub Fast Recompress
Constant quant 2 encode

Results : (sizes in Ko, min/avg/max I-frame in bytes)
Quant 2 :
clip 1) Full Motion=20
ME# | size | min/avg/max I-frames
1)    4302   48243/52388/57950
2)    4298   48246/52352/57381
2*)   4278   47713/52106/56938
3)    4384   48760/53421/58821
4)    4382   48904/53395/58786

2* : 2) with ADVANCEDDIAMOND8 instead of HALFPELDIAMOND8. No speed hit.

clip 2) Moves
ME# | size | min/avg/max I-frames
1)    4906    2139/8908/28966
2)    4900    2104/8896/28769
3)    4902    2073/8902/28855
4)    4902    2086/8900/28768

clip 3) Travellings/Zooms
ME# | size | min/avg/max I-frames
1)    5610    1333/10106/34231
2)    5600    1348/10089/33789
3)    5612    1360/10111/33935
4)    5608    1361/10103/33833

PSNR :
The PSNR results were always very close, less than 0,02 dB difference.

Speed :
PMVfast^2 and PMVfast advdiamond are the faster
mixed EPZS 5% slower
full EPZS 10% slower

Conclusion :

the best ME in term of compression is :
EPSZ_16 PMVfast_8
PMV_EARLYSTOP16 | PMV_HALFPELREFINE16 | PMV_EXTSEARCH16 | =
PMV_USESQUARES16 |
PMV_EARLYSTOP8 | PMV_HALFPELREFINE8 | PMV_ADVANCEDDIAMOND8

To avoid EPZS (buggy with luma masking??), best ME is :
PMVfast_16 PMVfast_8
PMV_EARLYSTOP16 | PMV_HALFPELREFINE16 | PMV_EXTSEARCH16 | =
PMV_ADVANCEDDIAMOND16 |
PMV_EARLYSTOP8 | PMV_HALFPELREFINE8 | PMV_ADVANCEDDIAMOND8

08.08.2002 @15h32GMT / by MarcFD / marc.fd#libertysurf.fr (#=3D@)
------=_NextPart_000_0023_01C23F90.789A16C0--