[M4IF Technotes] Rounding of motion vectors for chroma?

[Christoph Lampert]

On Thu, 7 Nov 2002, Mithun N wrote:
> hi..
> the motion vectors cant go to an accuracy of 1/32,
> the reason being that there can be max of 4 non
> transparent
> blocks in a macroblock, each having an accuracy of
> 1/4...which
> comes out to 1/16 and not 1/32.

Sorry to still bother you, but this is not the point. 
Since chroma is 1:2 _subsampled_, you have to divide 
chroma value not by K but by 2*K ! 
This _is_ clearly stated in the standard (right atop of the tables 7-6
to 7-9) and it's also completely logical. 
halfpel / (2*4)   is 16th-pel, okay, no problem. 
But for quarterpel this is _not_ working, because 
quarterpel / (2*4)  is 32th-pel ! 
I read the standard that this is not supposed to be done,
but that _before_ that quarterpel-values should be divided 
by 2 to obtain halfpel and _then_ add up and divide by 2*K. 
However there is no operator given for this dividing by 2. 
So should there be no rounding? And "no rounding" means 
round to -infinity? 
Then if one component in qpel would get values  
    3/4, 3/4, 1/4, 1/4  
the resulting vector would be (1+1+0+0)/2 /(4*2) ---Table_7.6---> 0/2 
I simply find this very strange and non-symmetric and would like to get
confirmation.
Christoph
> and as far as using the table is concerned the
> rounding is very
> clearly mentioned _how_ to be done.
> for eg..
> if a macroblock has 4 non transparent blocks,and
> you are calculating
> the chroma vector for the same then add the 4
> motion vectors(each with an accuracy of 1/4)
> Divide the sum by 4....the so obtained accuracy
> will be 1/16...
> Now check the table 7.6   which shows the rounding
> for 1/16 accuracy.
> If the fractional part of the motion vector is
> between 0/16 to 2/16 then the resulting fractional
> part becomes 0/2=0.0
> If the fractional part of the motion vector is
> between 3/16 to 13/16 then the resulting
> fractional part becomes 1/2=0.5
> If the fractional part of the motion vector is
> between 14/16 to 15/16 then the resulting
> fractional part becomes 2/2=1.0
> similarly if u have 3 nontransparent blocks ...you
> will have  amax accuracy of 1/4*1/3 =1/12..then
> check table 7.7 and perform similar steps as in
> table 7.6
> Do ask if still this doesnt solve your query
> regards
> Mithun Nayak
> 
> -----Original Message-----
> From: technotes-admin   lists.m4if.org
> [mailto:technotes-admin   lists.m4if.org]On Behalf
> Of Christoph Lampert
> Sent: Friday, November 01, 2023 6:26 PM
> To: technotes   lists.m4if.org
> Subject: RE: [M4IF Technotes] Rounding of motion
> vectors for chroma?
> 
> 
> Sorry, not really.
> 
> Of course I know that vectors have to be rounded
> and that table 7.6 to
> 7.9 should be used for that. But the exact method
> _how_ this is supposed
> to be done is very unclear.
> 
> If K=4 and the lumi motion vectors in quarterpel
> resolution,
> chroma (mv[0]+mv[1]+mv[2]+mv[3])/(2*K) would be
> 32th-pel(!)
> 
> This is not allowed (and there's no table for
> that), so the standard
> talks about dividing quarterpel values by 2 before
> doing the prodecure.
> (If I am not mistaken, this section has been
> changed between different
> versions of drafts etc.)
> So is rounding supposed to be a two-step process
> in quarterpel-inter4v
> mode? First devide by 2. Then use table 7-6? What
> rounding is used for the
> first division by 2?
> 
> Next, what about K=1? Then the result if mv[0]/2
> is 8th-pel.
> Should the 8th-pel table 7-8 be used for rounding
> as the reference
> software does (if I remember correctly)? Why isn't
> it two-step again?
> 
> And why is the whole process so complicated, why
> isn't there simply
> another table for 32th-pel to halfpel?
> 
> Christoph
> 
> 
> 
> 
> >
> > -----Original Message-----
> > From: technotes-admin   lists.m4if.org
> > [mailto:technotes-admin   lists.m4if.org]On Behalf
> > Of Christoph Lampert
> > Sent: Wednesday, October 30, 2023 8:21 PM
> > To: technotes   lists.m4if.org
> > Subject: [M4IF Technotes] Rounding of motion
> > vectors for chroma?
> >
> >
> > Hi,
> >
> > I have a question about rounding motion vectors,
> > especially for
> > chroma if quarterpel is active. I'll use a
> simple
> > example.
> > Consider a macro block in inter4v mode with
> > luminance motion vectors
> > mv[0] to mv[3] in halfpel or quarterpel
> > resolution.
> >
> > How do I calculate the motion vector for chroma
> > from them?
> > I know of course that it's
> > mv[0]+mv[1]+mv[2]+mv[3] / (2*4)
> > but how exactly is rounding supposed to be done?
> > What I read from the
> > standard about this is as unclear as possible.
> >
> > Thank you in advance,
> >
> > Christoph
> >
> > --
> > Christoph H. Lampert chl   math,uni-bonn,de |
> Diese
> > Signature wurde maschi-
> > Beringstr. 6, Zi. 15, 53115 Bonn, Germany | nell
> > erstellt und bedarf
> > Tel. (0228) 73-4708  Fax. +49 228 73-7916 |
> keiner
> > Unterschrift. AZ 27B-6
> >
> > _______________________________________________
> > Technotes mailing list
> > Technotes   lists.m4if.org
> > http://lists.m4if.org/mailman/listinfo/technotes
> >
> > _______________________________________________
> > Technotes mailing list
> > Technotes   lists.m4if.org
> > http://lists.m4if.org/mailman/listinfo/technotes
> >
> 
> _______________________________________________
> Technotes mailing list
> Technotes   lists.m4if.org
> http://lists.m4if.org/mailman/listinfo/technotes
> 
>