The discrepancies between PROF with affine AMVP and PROF with affine MERGE

[ZhipinDeng]

In section 8.5.6.3.1 of R2001-vA, whether a prediction block border extention is used for PROF is conditioned on “inter_affine_flag[ xSb ][ ySb ] && !ph_disable_prof_flag”, as shown below.

...
The prediction block border extension size brdExtSize is derived as follows:
brdExtSize = ( bdofFlag | | ( inter_affine_flag[ xSb ][ ySb ] && !ph_prof_disabled_flag ) ) ? 2 : 0 (934)
...
– The prediction luma sample value predSamplesLX[ xL ][ yL ] is derived as follows:

– If bdofFlag is equal to TRUE or ( ph_prof_disabled_flag is equal to FALSE and inter_affine_flag[ xSb ][ ySb ] is equal to TRUE ), ...

...

However, it causes undesirable different handling mechanisms between PROF with affine AMVP and PROF with affine MERGE, e.g., a prediction block of (sbWidth + 2) x (sbHeight +2) is used for PROF when PROF is appied to an affine AMVP block, but a prediction block of (sbWidth) x (sbHeight) is used for PROF when PROF is appied to an affine MERGE block.

It is suggested to correct the spec with the following modifications (highlighted as bold):

...
The prediction block border extension size brdExtSize is derived as follows:
brdExtSize = ( bdofFlag | | cbProfFlagLX ) ? 2 : 0 (934)
...
– The prediction luma sample value predSamplesLX[ xL ][ yL ] is derived as follows:

– If bdofFlag is equal to TRUE or cbProfFlagLX is equal to TRUE, ...

...

[hanhuang]

Should cbProfFlagLX be set as the input to 8.5.6.3?

[ZhipinDeng]

Yes, it should.

Replying to hanhuang:

Should cbProfFlagLX be set as the input to 8.5.6.3?

[bbross]

Thanks for brining that up.

Can experts familiar with this confirm that the proposed fix is aligned with VTM?

[bbross]

Fixed in JVET-S2001-v7 as agreed in the meeting.