Text/SW mismatch for MMVD scaling with ILRPs (8.2.5.7)

[fbossen]

There are conditions such as
– If RefPicList[ 0 ][ refIdxL0 ] is not an LTRP and RefPicList[ 1 ][ refIdxL1 ] is not an LTRP, the following applies

As far as I can tell, to match the behaviour of the software, these conditions should be
– If RefPicList[ 0 ][ refIdxL0 ] is not marked as "used for long-term reference" and RefPicList[ 1 ][ refIdxL1 ] is not marked as "used for long-term reference", the following applies

Furthermore, with such a modification, when only one of the two reference pictures is an ILRP, some odd scaling of MMVDs occurs.
When considering

mMvdL1[ 0 ] = Sign( currPocDiffL0 ) = = Sign( currPocDiffL1 ) ? mMvdL0[ 0 ] : −mMvdL0[ 0 ] (565)

the sign comparison always fails when one of the two reference pictures is an ILRP (because Sign(0) is 0), resulting in mMvdL1 (ILRP reference) being equal to -mMvdL0 (non-ILRP reference).

Should this be as follows? (even though the SW seems to match the text in this aspect)

mMvdL1[ 0 ] = Sign( currPocDiffL0 ) = = −Sign( currPocDiffL1 ) ? −mMvdL0[ 0 ] : mMvdL0[ 0 ] (565)

Changing the software to match the text could also be done and would result in the mMvd for an ILRP reference to always be 0 if the second reference is not an ILRP.

[jlchen]

Thanks for reporting the mismatch. The suggested changes of the first part make sense and will be fixed in JVET-S2001-vC.

I would suggest not making the second change since SW also need to be modified. I guess,for ILRP use case, MMVD wouldn't show obvious difference w/ or w/o the further modification.

[jlchen]

I'm closing this ticket. Please reopen it if the solution is not agreed.