[x264-devel] motion estimation

[Oleg Terner]

Hello,
We followed your advice and tried to understand the meaning of the 
mb.cache.mv vectors through
the simple examples. We've printed its values for run of a 2-frame 
movies with small number of blocks ( up to 4 ) each of them
contains only one constant number.
The values that we've recieved in mb.cache.mv aren't obvious ( here 
we're meaning
mb.cache.mv[ 0 ] that is used always ). Some of the vectors are zero, 
and some other have values that not always
meet with real motion ( sometimes it's halfpixel mistake and some times 
it's strongly different ) There is a me method influence also.

So can you, please, explain the structure of the mb.cache.mv? How and 
what vectors are stored in each of 48 vectors?

Thanks in advance,
Oleg.

Loren Merritt wrote:
> On Mon, 29 Sep 2008, Oleg Terner wrote:
>
>   
>> We just want to get the mv of each mb as compared to the previous frame.
>>
>> To make this simple, we took just 2 frames.
>> We see that the first frame is encoded as an I-frame, the second as a
>> P-frame.
>> However, the mv we got are sometimes very strange.
>>
>> The question is therefore as follows:
>>
>> How can we force the me routine to check mv only against the previous
>> frame.
>>     
>
> That's the default.
>
>   
>> and if so, where do we get these values from.
>>     
>
> h->mb.mv for the whole frame, if you access it after slice_write.
> h->mb.cache.mv for the current mb, if you access it after 
> macroblock_analyse.
> Both of the above are invalid for intra macroblocks.
> h->mb.mvr has mvs for every block for which motion search was run (which 
> is every block if you use --no-fast-pskip). These values are still 
> sometimes bogus; perhaps some intra blocks do have a reasonable mv 
> and intra was just better, but others are intra because motion 
> search couldn't find anything and just had to pick a random mv.
>
> Check your results against `mplayer -lavdopts vismv=1`.
>
> --Loren Merritt
> _______________________________________________
> x264-devel mailing list
> x264-devel at videolan.org
> http://mailman.videolan.org/listinfo/x264-devel
>
>