NumView derivation seems not correct. The current code recognized 6 views when there is only 3 views. I found this bug when decoding the bitstream in ​http://wftp3.itu.int/av-arch/jct3v-site/bitstream_exchange/under_test/MV-HEVC/MVHEVCS_I_Nokia_2.zip (but some additional workaround is needed for testing with the bistream)

The current code in getNumViews() [TComSlice.cpp]

Int lId = m_layerIdInNuh[i]; if( i > 0 && ( getViewIndex( lId ) != getScalabilityId( i - 1, VIEW_ORDER_INDEX ) ) ) {

numViews++;

}

should be replaced with

Int lId = m_layerIdInNuh[i]; if ( i > 0 ) {

Bool newViewFlag = true; for( Int j = 0; j < i; j++ ) {

if( getViewIndex( lId ) == getScalabilityId( getLayerIdInNuh(j), VIEW_ORDER_INDEX ) ) {

newViewFlag = false;

}

} if( newViewFlag ) {

numViews++;

}

}

As reported in the 3D-HEVC bug tracker at ​https://hevc.hhi.fraunhofer.de/trac/3d-hevc/ticket/106, the identical motion check may have problem in layered coding.

Specifically, in layerd coding, pictures has two dimentions, POC and layer, but we only check POCs in the identical motion check. Thus we may fail in the case that two reference pictures have the same POC but different layer id.

Bool TComPrediction::xCheckIdenticalMotion ( TComDataCU* pcCU, UInt PartAddr ) { ... #if SVC_EXTENSION

Int layerIdL0 = pcCU->getSlice()->getRefPic(REF_PIC_LIST_0, pcCU->getCUMvField(REF_PIC_LIST_0)->getRefIdx(PartAddr))->getLayerId(); Int layerIdL1 = pcCU->getSlice()->getRefPic(REF_PIC_LIST_1, pcCU->getCUMvField(REF_PIC_LIST_1)->getRefIdx(PartAddr))->getLayerId(); if(layerIdL0 == layerIdL1 && RefPOCL0 == RefPOCL1 && pcCU->getCUMvField(REF_PIC_LIST_0)->getMv(PartAddr) == pcCU->getCUMvField(REF_PIC_LIST_1)->getMv(PartAddr))

#else

if(RefPOCL0 == RefPOCL1 && pcCU->getCUMvField(REF_PIC_LIST_0)->getMv(PartAddr) == pcCU->getCUMvField(REF_PIC_LIST_1)->getMv(PartAddr))

#endif

Problem: In HEVC specification, picture type such as short term or long term is involved in the motion vector scaling. In HEVC, picture type of the reference pictures is defined at the time when the picture was current, particularly

The function LongTermRefPic( aPic, aPb, refIdx, LX ), with X being 0 or 1, is defined as follows: – If the picture with index refIdx from reference picture list LX of the slice containing prediction block aPb in the picture aPic was marked as "used for long term reference" at the time when aPic was the current picture, LongTermRefPic( aPic, aPb, refIdx, LX ) is equal to 1. – Otherwise, LongTermRefPic( aPic, aPb, refIdx, LX ) is equal to 0.

In SHVC, derived inter-layer reference picture can be used as a collocated picture and the picture type of its reference pictures can be involved in the motion vector scaling. However, since derived inter-layer reference picture is not a coded picture, it was not a current picture when LongTermRefPic is defined. So, the status of the derived inter-layer reference picture being current has to be defined for the marking purpose of its reference pictures to be complaint with the HEVC motion vector scaling process and “HLS” only concept.

Possible solution: The solution can be the following, by setting the resampled picture to be current during the motion field resampling (marked with bold): ... – When VpsInterLayerMotionPredictionEnabled[ LayerIdxInVps[ currLayerId ] ][ LayerIdxInVps[ rLId ] ] is equal to 1, the following steps apply: – The resampled picture rsPic is set to be the current picture at the time when the reference layer picture rlPic is the current picture – A single slice rsSlice of the resampled picture rsPic is generated as follows: ...