[x264-devel] Re: Regarding quarter pixel ME
Radek Czyz
radoslaw at syskin.cjb.net
Thu Mar 29 09:50:52 CEST 2007
Well the same thing happens now as it did when you refined
fullpel->hpel. One of the eight qpixel neighbours might give you better
SAD in which case you pick it. If it was one of the hpels that gives you
better SAD, pick hpel.
mahesh k wrote:
> Dear,
>
> What i mean was,,,First I calculated integer level best motion vector
> using SAD as error criterion. that is the one (block) which gives
> minimum SAD . In this blk size is 16 x 16 and reference frame is
> previous frame(non interpolated)
>
> Then surrounding best integer level motion vector I choosed the eight
> half pixel locations and searched for the besxt match among this using
> SAD as the criterion. In this case also blk size was 16 x16. Now I
> compared the minimum SAD in both cases and if half pixel min SAD is best
> then I will go for quarter pixel level....In half and quarter pixel
> level the blk size was 16 x16 but the reference frame was previous frame
> interpolated upto quarter pixel level.
>
> Ex:
>
> curr blk to be matched (2x2)
>
> 100 102
>
> 90 95
>
> Reference (non interpolated):
>
> 80 83 95 100
>
> 100 40 60 75
>
> 120 60 80 90
>
> 90 50 70 88
>
>
> Reference (non interpolated) used as reference for integer pixel ME.
>
> 80 83 95 100
>
> 100 40 60 75
>
> 120 60 80 90
>
> 90 50 70 88
>
> Reference (interpolated) frame used as reference for half and quarter
> pixel ME.
>
> I Q H Q I Q H Q I
> 80 82 81 82 83 86 89 85 95 93 97 98 100
>
> 100 85 70 55 40 45 50 55 60 64 68 72 75
>
> 120 105 90 75 60 65 70 75 80 83 84 88 90
>
> 90 60 70 40 50 55 60 65 70 75 79 84 88
>
> This is how i implemented ....
>
> Regards,
> Maheshwar.
>
>
> On 3/29/07, *Radek Czyz* <radoslaw at syskin.cjb.net
> <mailto:radoslaw at syskin.cjb.net>> wrote:
>
> Hi, can you explain more clearly? It's not ME that gives you SAD - it's
> SAD that can be used to perform ME. The other way around ~
>
> mahesh k wrote:
> > Ok, thx for ur reply.....Can u plz tell me whether the quarter
> pixel ME
> > will give the min SAD for all the macroblocks within a frame.?
> Also plz
> > tell me where I can find the test results regarding this?
> >
> > regards,
> > mahesh
> >
> > On 3/28/07, *Radek Czyz* < radoslaw at syskin.cjb.net
> <mailto:radoslaw at syskin.cjb.net>
> > <mailto:radoslaw at syskin.cjb.net
> <mailto:radoslaw at syskin.cjb.net>>> wrote:
> >
> > > U plz tell me whether this was correct or not.
> >
> > Well yeah, it's a good (although wasteful in terms of memory)
> way of
> > doing it.
> >
> > You ask for "correct" but ME is never defined, so all methods are
> > "correct". Just some are more efficient (give better or
> faster result)
> > than others.
> >
> > mahesh k wrote:
> > > I went through the JM reference code but I was not able to
> understand
> > > it. Recently i implemented quarter pixel motion
> estimation. First I
> > > interpolated the whole reference frame upto quarter pixel
> level and
> > > stored this in a buffer. Then I found the integer pixel
> motion
> > vector
> > > using the previos frame as reference frame (non interpolated
> > reference
> > > frame) and the search criterion i used was SAD. Then I
> performed
> > the sub
> > > pixel refinement using the best integer pixel motion vector
> > coordinates
> > > as center using the quarterly interpolated frame as reference.
> > >
> > > U plz tell me whether this was correct or not.
> > >
> > > Regards,
> > > Mahesh.
> > >
> > > On 3/28/07, *Tomas Carnecky* <tom at dbservice.com
> <mailto:tom at dbservice.com>
> > <mailto:tom at dbservice.com <mailto:tom at dbservice.com>>
> > > <mailto: tom at dbservice.com <mailto:tom at dbservice.com>
> <mailto:tom at dbservice.com <mailto:tom at dbservice.com>>>> wrote:
> > >
> > > mahesh k wrote:
> > > > I didnot get u exactly..Can u tell me in detail...
> > > >
> > >
> > > RTFM means 'Read The Fucking Manual', see[1]. He
> replaced the
> > 'M' with a
> > > 'S' which he afterwards explained to be, quote,
> 'sourcecode or
> > > standard'. That means, he was trying to tell you 'Read
> The F***
> > > sourcecode' or Read The F*** standard'.
> > > IOW (you can use wikipedia, too, to look up what that
> > means), he told
> > > you to read the sourcecode (which is freely available)
> or the
> > H.264
> > > standard. The information you're looking for is in both
> > locations,
> > > hence, quote, 'your choice'.
> > >
> > > >
> > > > On 3/27/07, *Loren Merritt*
> <lorenm at u.washington.edu <mailto:lorenm at u.washington.edu>
> > <mailto:lorenm at u.washington.edu <mailto:lorenm at u.washington.edu>>
> > > <mailto: lorenm at u.washington.edu
> <mailto:lorenm at u.washington.edu>
> > <mailto: lorenm at u.washington.edu
> <mailto:lorenm at u.washington.edu>>>
> > > > <mailto: lorenm at u.washington.edu
> <mailto:lorenm at u.washington.edu>
> > <mailto: lorenm at u.washington.edu
> <mailto:lorenm at u.washington.edu>>
> > > <mailto:lorenm at u.washington.edu
> <mailto:lorenm at u.washington.edu>
> > <mailto:lorenm at u.washington.edu
> <mailto:lorenm at u.washington.edu>> >>> wrote:
> > > >
> > > > On Tue, 27 Mar 2007, mahesh k wrote:
> > > >
> > > > > Can any one plz tell me how to choose the
> refernece
> > frame for
> > > > > integer level and quarter pixel motion
> estimation.
> > > >
> > > > RTFS. (sourcecode or standard, your choice)
> > > >
> > >
> > > tom
> > >
> > > [1] http://en.wikipedia.org/wiki/RTFM
> > < http://en.wikipedia.org/wiki/RTFM>
> > >
> > > --
> > > This is the x264-devel mailing-list
> > > To unsubscribe, go to:
> http://developers.videolan.org/lists.html
> > >
> > >
> >
> > --
> > This is the x264-devel mailing-list
> > To unsubscribe, go to: http://developers.videolan.org/lists.html
> >
> >
>
> --
> This is the x264-devel mailing-list
> To unsubscribe, go to: http://developers.videolan.org/lists.html
> <http://developers.videolan.org/lists.html>
>
>
--
This is the x264-devel mailing-list
To unsubscribe, go to: http://developers.videolan.org/lists.html
More information about the x264-devel
mailing list