Mismatch between spec and VTM in constructed affine candidate

[emora]

There is a mismatch between the spec and VTM software concerning the derivation of the last constructed affine control point motion vector merge candidate (Const6). Equation 824 states:

cpMvLXCorner[ 1 ][ 1 ] = (cpMvLXCorner[ 0 ][ 1 ] << 7) + ((cpMvLXCorner[ 2 ][ 0 ] − cpMvLXCorner[ 0 ][ 0 ]) << (7 + Log2(cbWidth) − Log2(cbHeight)))

where as the computation done in VTM is as follows:
vy = (cMv[l][0].ver << shift) - ((cMv[l][2].hor - cMv[l][0].hor) << shiftHtoW);

There seems to be a substraction in VTM whereas it is an addition in the spec. I believe the spec should be modified to match the software in this case. JVET-K1001 v6 had the correct formula.

[bbross]

Thanks for reporting.

This has been fixed in JVET-Q2001-v8.