A bug is found in HTM10.0rc1. Generally speaking, the original code tries to use InterDir to determine the prediction direction. However, InterDir is not compressed as MV and RefIdx. Thus a misalignment may occur. In some worse case, the codec may crash, although that does not happen in CTC. Attached please find a possible fix with macro HTM10RC1_FIX and the simulation results under CTC. There is a very minor difference (<0.001%) in the last column.

At line 4746 in TComDataCU.cpp, it should be #if !HTM10RC1_FIX

if( ucInterDir & 1 ) {

#endif

pcTextureCU->getMvField( pcTextureCU, uiPartIdx, REF_PIC_LIST_0, cMVField );

#if HTM10RC1_FIX

if( cMVField.getRefIdx() >= 0 ) {

#endif ...

At line 4775 in TComDataCU.cpp, it should be

#if !HTM10RC1_FIX

if( !tmpDir && ( ucInterDir & 2 )) {

#endif

pcTextureCU->getMvField( pcTextureCU, uiPartIdx, REF_PIC_LIST_1, cMVField );

#if HTM10RC1_FIX

if( !tmpDir && cMVField.getRefIdx() >= 0 ) {

#endif

Thank you!

In HTM 9.3, the wrong motion information is used for pruning of DV candidate.

Line 4497 in TComDataCU.cpp is if(puhInterDirNeighbours[iCount] == puhInterDirNeighbours[iCnloop] && pcMvFieldNeighbours[iCnloop<<1]==pcMvFieldNeighbours[(iCount<<1)] && pcMvFieldNeighbours[(iCnloop<<1)+1]==pcMvFieldNeighbours[(iCount<<1)+1])

It should be fixed to if(ivCandDir[1] == puhInterDirNeighbours[iCnloop] && pcMvFieldNeighbours[iCnloop<<1]==tmpMV[0] && pcMvFieldNeighbours[(iCnloop<<1)+1]==tmpMV[1])

In the current HTM, motion/disparity vectors are clipped before storing them in some cases although they should be clipped before generating prediction signals. In other words, motion/disparity vectors in the motion storage should not be clipped.

At least, I found such wrong clippings in TComDataCU::xAddVspCand, TComDataCU::xAddIvMRGCand, TComDataCU::xGetInterMergeCandidates (6 places), TComDataCU::getDisMvpCandNBDV (2 places), TComDataCU::estimateDVFromDM, TComDataCU::xCheckSpatialNBDV, TComDataCU::getInterViewMergeCands (3 places). After removing these clipping, it becomes necessary to add clipping in TComPrediction::xPredInterUniARP since NBDV becomes a non-clipped vector.