[M4IF Technotes] Is OBMC implementation different from Momusys and 14496-2??

[Luca Piccarreta]

I'm also very interested in this point.
The spec for OBMC look very encoder-oriented.
Encoders usually have all the MVs available at
encoding time.
Decoders, instead, work in a pipeline.
In MS software also bottom vector is not used.
And having the right vector available really
complicates decoding.
In fact, Motion compensation for a MB is carried
out only after the subsequent MB header and MV
(and shape) have been decoded.
This may be (Not sure, though) the reason why
OBMC is not present in any profile/level
combination.
So one may wonder why there is a tool that no encoder
can use being sure that any decoder will ever be able
to decode!
Strictly following the specs for OBMC would heavily
impact on performances (DCT coefs or output) for a whole
row should be cached.
Partially following the specs makes the decoder code
cumbersome to seay the least (see decode_PVOP...
code in MS Ref).
Thanks to anyone that will shed light on this issue.
Luca Piccarreta.
----- Original Message -----
From: <CC_Ju   mtk.com.tw>
To: <technotes   lists.m4if.org>
Sent: Thursday, October 17, 2023 9:08 AM
Subject: [M4IF Technotes] Is OBMC implementation different from Momusys and
14496-2??
> Dear all:
>     I don't know if I misunderstand the OBMC implementation in Momusys
> software. It seems to use only up, left, right, and current macroblock's
> motion vector to do OBMC operation. The motion vector in bottom macroblock
> seems don't used in OBMC operation. This differs from 14496-2
> specification. Anybody can help?
>
> Thanks in advance.
>
> ---
> Chi-Cheng
>
>
>
>
>
>
>
>
> _______________________________________________
> Technotes mailing list
> Technotes   lists.m4if.org
> http://lists.m4if.org/mailman/listinfo/technotes