[Multilayer Profiles] Potential Mismatch of VTM22.0 & Specification Related To Derivation Process For Merge Motion Vector Difference

[ksthey]

In "8.5.2.7 Derivation process for merge motion vector difference", Equation 563:
mMvdL1[ 0 ] = Sign( currPocDiffL0 ) == Sign( currPocDiffL1 ) ? mMvdL0[ 0 ] : −mMvdL0[ 0 ].

The equivalent code is in VTM22.0, InterPrediction.cpp, MergeCtx::getMmvdDeltaMv(), starting from line 2478:

if ((poc1 - currPoc)*(poc0 - currPoc) > 0)
{

deltaMv[1] = deltaMv[0];

}
else
{

deltaMv[1].set(-1 * deltaMv[0].getHor(), -1 * deltaMv[0].getVer());

}

According to specification, definition of Sign(x):
Sign(x) = 1; x > 0
Sign(x) = 0; x == 0
Sign(x) = -1; x < 0

Let's assume:

• result of condition below in specification as r_spec: r_spec = Sign( currPocDiffL0 ) == Sign( currPocDiffL1 )
• result of condition below in VTM22.0 as r_vtm: r_vtm = (currPocDiffL1 * currPocDiffL0 > 0)

Use combination of currPocDiffL0 & currPocDiffL1 as input:
1.) (currPocDiffL0 < 0) && (currPocDiffL1 < 0)

r_spec = 1; r_vtm = 1; ==> r_spec == r_vtm;

2.) (currPocDiffL0 < 0) && (currPocDiffL1 > 0)

r_spec = 0; r_vtm = 0; ==> r_spec == r_vtm;

3.) (currPocDiffL0 > 0) && (currPocDiffL1 < 0)

r_spec = 0; r_vtm = 0; ==> r_spec == r_vtm;

4.) (currPocDiffL0 < 0) && (currPocDiffL1 < 0)

r_spec = 1; r_vtm = 1; ==> r_spec == r_vtm;

5.) (currPocDiffL0 == 0) && (currPocDiffL1 != 0)

r_spec = 1; r_vtm = 1; ==> r_spec == r_vtm;

6.) (currPocDiffL0 != 0) && (currPocDiffL1 == 0)

r_spec = 1; r_vtm = 1; ==> r_spec == r_vtm;

7.) (currPocDiffL0 == 0) && (currPocDiffL1 == 0)

r_spec = 1; r_vtm = 0; ==> r_spec != r_vtm;

Case 5, 6 & 7 are only possible in multilayer profile (current picture and reference picture same POC, but different layers).
As shown above, for case 7, results of specification & VTM are different.

So, which is the correct one? specification or VTM?

[bbross]

Thanks for reporting and I see the mismatch for the multi-layer case. This should be fixed in v4 but before, we need to clarify which one is the intended behaviour.

[fbossen]

AFAICT, this came from proposal JVET-M0332, which included the following draft text (consistent with the software):

Otherwise, the following applies:
mMvdL0[ 0 ] = (currPocDiffL0*currPocDiffL1>0)?mMvdL1[ 0 ]: -mMvdL1[ 0 ];
mMvdL0[ 1 ] = (currPocDiffL0*currPocDiffL1>0)?mMvdL1[ 1 ]: -mMvdL1[ 1 ];

It was integrated in JVET-N1001v3 as:

Otherwise, the following applies:
mMvdL10[ 0 ] = Sign( currPocDiffL0 ) == Sign( currPocDiffL1 ) ? mMvdL1[ 0 ] : −mMvdL1[ 0 ] (8 375)
mMvdL0[ 1 ] = Sign( currPocDiffL0 ) == Sign( currPocDiffL1) ? mMvdL1[ 1 ] : −mMvdL1[ 1 ] (8 376)

Most likely, implications for multilayer were overlooked at the time and it would be difficult to claim that a particular behaviour was intended.

This issue needs to be resolved to address a ballot comment on VVC reference software.

[hmjang]

Hi all,
I want to understand the issue on VTM regarding this.
the claimed case is the CASE7 ( currPocL0 == 0 and CurrPocL1 == 0).

at least this case looks like being gated by
"If currPocDiffL0 is equal to currPocDiffL1, the following applies" which is equation (548) above in the SPEC.
mMvdL0[ 0 ] = MmvdOffset[ xCb ][ yCb ][ 0 ] (548)
mMvdL0[ 1 ] = MmvdOffset[ xCb ][ yCb ][ 1 ] (549)
mMvdL1[ 0 ] = MmvdOffset[ xCb ][ yCb ][ 0 ] (550)
mMvdL1[ 1 ] = MmvdOffset[ xCb ][ yCb ][ 1 ] (551)

and also in the VTM, this part is same as shown below:

if ((poc0 - currPoc) == (poc1 - currPoc))
{

deltaMv[1] = deltaMv[0];

}

Do I miss something to understand issue?

[fbossen]

@hmjang Thank you for pointing out the gating condition. I had missed that. So it looks like there is no mismatch after all because the "mismatched" condition is never reached in case 7.