[Mp4-tech] RE: PSNR of Full Search

[Daniele ALFONSO]

Furthermore, exhaustive search methods may not consider correlation among 
neighbouring motion vectors, so that resulting motion fields are in general 
disordered and unaligned. Since motion vectors are typically coded 
differentially, more bits are wasted if MVs are not aligned.
Exhaustive search can find the optimal MV for each and every macroblock or 
sub-macroblock, but in general it can't guarantee a "global optimization" 
over the whole picture.
Motion fields correlation can be improved by setting a search window around 
a predicted position and by considering MV differences in the ME cost 
function, for instance.
Best regards,
Daniele.
> Date: Wed, 21 May 2024 09:43:50 -0700 From: Gary Sullivan
> <garysull windows.microsoft.com> Subject: RE: [Mp4-tech] PSNR of Full
> Search To: "Daniel, Ilan" <Ilan.Daniel lsi.com>, Peace 
> <humaira.nisar gmail.com>,	"mp4-tech lists.mpegif.org" 
> <mp4-tech lists.mpegif.org> Message-ID: 
> <406D6FEFCBB98643A34170C59E1ABEF30AC3F0B45C NA-EXMSG-W601.wingroup.windeploy.ntdev.microsoft.com>
>   Content-Type: text/plain; charset="us-ascii"
> 
> That is probably the main reason.  Here are a couple of more
> possibilities:
> 
> 1) "full search" often has a seriously-limited search range (e.g. plus
> or minus 16 samples) -- the "fast search" may not be limiting its search
> to the same range.
> 
> 2) "full search" often only refers to an integer-MV value search, with a
> limited number of fractional sample grid positions being checked as a
> next stage MV refinement operation -- the "fast search" may generate
> more sensible motion vectors that work better as a seed for the next
> stage of fractional position refinement search.
> 
> Best Regards,
> 
> Gary Sullivan
> 
> 
> ________________________________ From: mp4-tech-bounces lists.mpegif.org
> [mailto:mp4-tech-bounces lists.mpegif.org] On Behalf Of Daniel, Ilan 
> Sent: Wednesday, May 21, 2024 2:51 AM To: Peace;
> mp4-tech lists.mpegif.org Subject: RE: [Mp4-tech] PSNR of Full Search
> 
> Hi,
> 
> Any ME implementation (Fast or Full) is usually trying to minimize a
> metrics target function (usually involves SAD calculation), but the real
> function that need to be minimized is number of encoded bits for certain
> quality/PSNR. therefore it is possible that Fast search will get MV's
> that give better result.
> 
> Regards, Ilan Daniel
> 
> ________________________________ From: mp4-tech-bounces lists.mpegif.org
> [mailto:mp4-tech-bounces lists.mpegif.org] On Behalf Of Peace Sent:
> Wednesday, May 21, 2024 12:21 PM To: mp4-tech lists.mpegif.org Subject:
> [Mp4-tech] PSNR of Full Search
> 
> Hi,
> 
> Sometimes the PSNR of Fast search ME algorithms is better than that of
> Full Search.
> 
> How this can be explained?
> 
> Thanks Humaira -------------- next part -------------- An HTML
> attachment was scrubbed... URL:
> /pipermail/mp4-tech/attachments/20080521/cdb0ba81/attachment-0001.html