FW: [Mp4-tech] Motion Compensation Quarter Pixel Interpolation

[Tung yi-Shin]

Hi, 
In my knowledge, there was some inconsistency about description and 
implementation of quarter-pel motion compensation between spec and refsoft. 
The problem was reported to video committee for over one year. After 
discussion by the committee, we decided to follow the reference software due 
to the following reasons.
1. most programmers realized their codec by referring to the software.
2. the conformance bitstreams are mostly generated by this way.
3. the adopted one processes horizontal filtering first then followed by 
vertical filtering. This way is logically simple and more regular, compared 
to what is originally specified by text or now adopted by AVC.
As a result, video committee issued the V2/COR2 and V3/COR1 to make this 
aspect more clear. We believe now the text and software are consistent.
As to the implementations by other public or commercial softwares, they 
should be changed accordingly, e.g., divx fixed its implementation in 
version 5.0.3.
what effect the two different implementaions may have? This is an 
interesting question. I have no idea about how ffmpeg does now, but I 
believe the difference between the two is not too much, because both methods 
will use the same 8x8 pixels to generate a predicted sample. If any one has 
more experiences of this, pls kindly make your comment.
Best Regards,
Yi-Shin
On Wed, 24 Mar 2024 23:20:40 +0800, Tihao Chiang wrote
> Let's talk about this.
> 
> Tihao
> 
> Tihao Chiang, Associate Professor,
> Dept. of Electronics Engineering, National Chiao-Tung University,
>  1001 Ta Hsueh Rd., Hsinchu, 30010, Taiwan, R.O.C. Email: 
tchiang mail.nctu.edu.tw
> Tel  :+886-351-31558 Fax  :+886-357-31791
> ----------------------
> ½±­}»¨,¥æ¤j¹q¤l©Ò
> ·s¦Ë¥«¤j¾Ç¸ô1001¸¹
> ----------------------
> 
> -----Original Message-----
> From: Gary Sullivan [mailto:garysull windows.microsoft.com] 
> Sent: Wednesday, March 24, 2024 3:03 AM
> To: ohm ient.rwth-aachen.de; Tihao Chiang
> Subject: FW: [Mp4-tech] Motion Compensation Quarter Pixel Interpolation
> 
> Isn't this something the committee reviewed and made a decision about?
> 
> -Gary
> 
> -----Original Message-----
> From: mp4-tech-bounces lists.mpegif.org
> [mailto:mp4-tech-bounces lists.mpegif.org] On Behalf Of Piyush Kaul
> Sent: Tuesday, March 23, 2024 6:02 AM
> To: Erik Schlanger
> Cc: mp4-tech lists.mpegif.org
> Subject: Re: [Mp4-tech] Motion Compensation Quarter Pixel Interpolation
> 
> Seems to be a problem to me too.
> As I understand the specs , only full-pel or other half-pel locations
> should be used for interpolating  half-pel positions. The quarter-pel
> positions are
> then obtained by bilenear interpolation using the half-pel and 
> quarter pel positions . The microsoft decoder seems to do it in the 
> wrong order. The values for any combination of XFrac=1,3 and YFrac=1,
> 2,3 should be wrong.
> 
> Regards
> Piyush
> 
> Erik Schlanger wrote:
> 
> > I am looking at the way quarter pixel interpolation is performed for
> > Mpeg4 motion compensation. The way the interpolation is performed in
> > ffmpeg is different than the way it is done by the ISO MPEG4 reference
> > decoder. Specifically, the ffmpeg code will do all neccesary 8-tap
> > filtering prior to doing any bilinear interpolation. However, in the
> ISO
> > MPEG4 reference decoder all 8-tap filtering for the horizontal
> direction
> > will be performed followed by bilinear interpolation in the horizontal
> > direction. The results of this are then fed to the vertical 8-tap
> > filter, followed by the vertical bilinear filter.
> >
> > Can anyone comment on what effect this different implementaion may
> have?
> >   I have consulted the specification and it is unclear to me. Looking
> at
> > numerical outputs the two implementations do produce different
> results.
> >
> > Thanks,
> > Erik
> >
> > _______________________________________________
> > Mp4-tech mailing list
> > Mp4-tech lists.mpegif.org
> > http://lists.mpegif.org/mailman/listinfo/mp4-tech
> >
> > Note: Conduct on the mailing list is subject to the Antitrust
> guidelines found at
> http://www.mpegif.org/public/documents/vault/mp-out-30042-Antitrust.php

--
Communication and Multimedia Laboratory
Dept. of Computer Science and Information Engineering, NTU